Journal of Urban Economics 108 (2018) 124–140

Contents lists available at ScienceDirect

Journal of Urban Economics

journal homepage: www.elsevier.com/locate/jue

The mobility of displaced workers: How the local industry mix affects job
search
Frank M.H. Neffke a,∗, Anne Otto b, César Hidalgo c
a
Center for International Development at Harvard University, 79, J.F.K. Street, MA, Cambridge, 02138, USA
b
Institutes for Employment Research (IAB), United States
c
Massachusetts Institute of Technology, The MIT Media Lab, United States

a r t i c l e i n f o a b s t r a c t

JEL classification: Are there Marshallian externalities in job search? We study how workers who lose their jobs in establishment
J24/J61/J64/R12, closures in Germany cope with their loss of employment. About a fifth of these displaced workers do not return
to social-security covered employment within the next three years. Among those who do get re-employed, about
Keywords:
two-thirds leave their old industry and one-third move out of their region. However, which of these two types
Displacement
Agglomeration externalities of mobility responses workers will choose depends on the local industry mix in ways that are suggestive of
Matching Marshallian benefits to job search. In particular, large concentrations of one’s old industry makes it easier to
Mobility find new jobs: in regions where the pre-displacement industry is large, displaced workers suffer relatively small
earnings losses and find new work faster. In contrast, large local industries skill-related to the pre-displacement
industry increase earnings losses but also protect against long-term unemployment. Analyzed through the lens
of a job-search model, the exact spatial and industrial job-switching patterns reveal that workers take these
Marshallian externalities into account when deciding how to allocate search efforts among industries.

Employers are apt to resort to any place where they are likely to
find a good choice of workers with the special skill which they re-
quire; while men seeking employment naturally go to places where
there are many employers who need such skill as theirs and where
therefore it is likely to find a good market. The owner of an isolated
factory, even if he has access to a plentiful supply of general labour,
is often put to great shifts for want of some special skilled labour;
and a skilled workman, when thrown out of employment in it, has
no easy refuge. (Marshall, 1890, IV.X.9).
1. Introduction
Marshallian externalities, i.e., benefits afforded by dense concentra-
tions of firms in the same economic activity, are sometimes associated
with the thickness of local labor markets. Traditionally, the importance
of local labor markets has been attributed to two separate, yet related,
mechanisms. First, firms benefit from locating close to other firms in
their industry as it would help them find workers with specialist skills.
Second, specialist workers are attracted to such geographical clusters,
because, if they were to lose their job, a local concentration of employ-
ers in their industry would make it easier to find new work that matches
their skill profiles. However, in spite of ample research on Marshallian
benefits that accrue to firms, the (re)employment benefits for workers
have received comparatively little attention in the urban economics lit-
erature. In this paper, we aim to shed light on the existence of Marshal-
lian externalities in job search by studying the careers of workers who
lose their jobs when establishments close down. We pose two questions.
First, do the career consequences of job displacement depend on the ex-
act mix of industries that exists in a local economy? We show that Mar-
shallian externalities manifest themselves in high re-employment rates
and low wage losses. And second, are workers aware of the existence of
such Marshallian externalities when they decide how to allocate search
efforts among different industries? Building on a model of job search, we
find that this is indeed the case: spatial patterns of new job matches sug-
gest that workers adjust their search strategies in ways that would allow
them to take better advantage of Marshallian job-search externalities.
In answering these questions, the paper connects debates on agglom-
eration economies to a large and growing literature in labor economics
that focuses on workers who lose their jobs in establishment closures.
Studying samples of these so-called “displaced” workers is attractive be-
cause establishment closures leave workers looking for new jobs when
they neither planned on, nor contributed to, the termination of their
current employment. As a consequence, such workers are relatively un-
affected by the self-selection problems that arise when job loss is an
endogenous outcome of interactions between workers and their employ-

∗
Corresponding author.
E-mail addresses: frank_neffke@hks.harvard.edu (F.M.H. Neffke), anne.Otto@iab.de (A. Otto), hidalgo@mit.edu (C. Hidalgo).

https://doi.org/10.1016/j.jue.2018.09.006
Received 20 January 2017; Received in revised form 26 September 2018
Available online 18 October 2018
0094-1190/© 2018 Elsevier Inc. All rights reserved.
F.M.H. Neffke et al.
ers. The literature on establishment closures has amply documented how
detrimental job displacement can be to people’s careers and well-being,
with consequences ranging from wage reductions and un- or underem-
ployment, to physical and mental health issues.
However, although displacement-related job-loss itself may plausibly
be exogenous to workers’ career plans, their response to it isn’t. After
all, workers can deal with unemployment in several ways. For instance,
they can search for jobs in their old industry or try their luck in another
industry. Similarly, workers can search for local jobs or move to other
regions. Which strategy they choose, and the likelihood of its success,
will depend on the kind of jobs a region has to offer. Consequently, the
rate at which workers change industries or move to other regions or
both, as well as the time it takes to find new jobs, will depend on which
jobs currently exist in the region. That is, they will depend on the local
mix of industries.
The importance of local labor markets is widely acknowledged in the
urban economics literature that deals with agglomeration externalities,
both in theoretical models and in empirical studies. For instance, Helsley
and Strange (1990) develop a model in which larger cities allow for a
better fit between a worker’s skill endowments and an employer’s skill
requirements. Such matching externalities are not limited to the quality
of the match. Duranton and Puga (2004) show that in matching models
with search frictions also the probability of successful matches will in-
crease as cities grow larger. That is, as the pools of potential employers
and employees grow, the likelihood that a worker remains unemployed
goes down. Indirect empirical evidence for such Marshallian labor mar-
ket pooling benefits have been found in local industries’ growth rates
(e.g., Dauth, 2010) and productivity (e.g., Feser, 2002), as well as in
industries’ overall spatial concentration (Rosenthal and Strange, 2001)
and coagglomeration patterns (Ellison et al., 2010). However, relatively
little is known about the validity of Marshall’s (1890) original claim that
industrial agglomerations help unemployed workers find new jobs. In
particular, we have incomplete answers to questions such as: Do unem-
ployed workers find jobs faster in regions with large local concentrations
of their prior industry? Do they suffer smaller wage losses in such re-
gions? And, finally, does the local industry mix affect how workers cope
with unemployment in terms of industrial or geographical mobility?
The first contribution of this paper is to provide answers that ad-
dress the causal mechanisms these questions imply by exploiting data
on displaced workers. To do so, we apply a combination of matching
techniques and regression models to a dataset that covers the employ-
ment history of over 20 million workers in Germany. Using difference-
in-differences estimation, we first show the causal effects of job displace-
ment on post-displacement wages and careers. We find that workers who
are displaced in establishment closures are not only less likely to return
to jobs covered by social security and more likely to experience sig-
nificant earnings losses. Those who do return to such jobs are also 66
percentage points (pp) more likely to change industries and 33 pp more
likely to change regions than their statistical twins. Next, we show that
there is substantial heterogeneity in these displacement effects that can
be attributed to Marshallian externalities. First, we find that a strong lo-
cal presence of the pre-displacement industry – i.e., the industry’s share
of regional employment is in the top instead of the bottom third of all
local industries – reduces post-displacement industry and region switch-
ing rates by 31 percent, respectively 12 percent. In contrast, high shares
of local employment in industries related to the pre-displacement indus-
try increase industry switching rates substantially, but do not prevent
workers from leaving the region. Turning to wages and re-employment
rates, we find substantial moderating effects of the local industry mix.
Whereas, on average, earnings drop by 39%, this drop is reduced to 32%
in regions with large concentrations of the pre-displacement industry.
Moreover, with 24% and 7% lower long-run nonemployment rates, hav-
ing high instead of low concentrations of the pre-displacement and re-
lated industries in the region offers some protection against long-term
nonemployment.
125
Journal of Urban Economics 108 (2018) 124–140
Having shown that Marshallian effects play a role in whether and
where unemployed workers find new jobs, next we ask: Do job searchers
take advantage of such Marshallian externalities? To provide a frame-
work for answering this question, we build on a search model by Fallick
(1992, 1993) in which unemployed workers divide their search efforts
between two sectors: their own industry and a sector composed of suit-
able alternative (i.e., related) industries. We repurpose this model for the
above question by assuming that greater search efforts translate into a
widening of the geographical search radius.
As a consequence, we can learn about shifts in the (unobserved) al-
location of search effort between the two sectors by looking at the geo-
graphical mobility of workers. The model predicts that favorable local
conditions in a particular sector not only increase the likelihood of find-
ing a job in this sector, but also induce workers to allocate more of their
search efforts to this sector, at the expense of the other sector. As a
consequence, favorable conditions in one sector will reduce the spatial
scope of search in the other sector. This prediction finds strong support
in the data.
Through our findings, we contribute to the debate on industrial spe-
cialization and diversity in regions (e.g., Glaeser et al., 1992; Henderson
et al., 1995; Porter, 2003) by placing the issue of agglomeration exter-
nalities in the context of job displacement. Our findings also shed light
on the importance of inter-industry relatedness, a topic of increasing in-
terest in economic geography (Delgado et al., 2010; Ellison et al., 2010;
Florida et al., 2011; Diodato et al., 2018). In particular, the finding that
skill-related employment induces workers to change industries, while
decreasing the likelihood of protracted nonemployment spells, shows
that clusters of related activities not only create agglomeration external-
ities for firms (Delgado et al., 2010; Neffke et al., 2012) but also for local
workers. Finally, our finding that workers take Marshallian externalities
into account when deciding how to divide search efforts provides fur-
ther (and more robust) support beyond Fallick (1993) for the existence
of strategic search as posited in wage search theory (e.g., Mortensen,
1986).
2. Literature review
Establishment closures have a profound impact on workers’ lives (see
Carrington and Fallick (2015) for a recent review). Apart from pecu-
niary losses, displaced workers suffer increased addiction problems and
a deterioration of their health (Black et al., 2015; Eliason and Storrie,
2009). Income losses after displacement can be severe and long-lived,
depressing incomes for periods of ten years or longer (e.g. Jacobson
et al., 1993; Couch and Placzek, 2010; Davis and von Wachter, 2011).
These income losses have been attributed to the loss of firm-specific
human capital (Becker, 1962), of back-loaded wage payments designed
to disincentivize shirking (Lazear, 1979), and of the “match capital”
(Jacobson et al., 1993, p. 686) workers have accumulated through find-
ing progressively better matching jobs over the course of their careers.
Earnings losses can materialize through protracted unemployment
spells and a reduction in daily wages (Carrington and Fallick, 2015).
In Germany, a major factor contributing to the loss of earnings is un-
employment (Burda and Mertens, 2001; Nedelkoska et al., 2015), es-
pecially immediately following displacement (Schmieder et al., 2010).
What determines how quickly displaced workers find new jobs? Previ-
ous research has pointed to national-level economic conditions: adverse
effects of displacement are more severe in macroeconomic downturns
(Davis and von Wachter, 2011) and in declining industries (Howland
and Peterson, 1988; Fallick, 1993). However, also local economic condi-
tions matter. First, the size and growth of local economies will affect the
arrival rate and the distribution of wage offers, both of which determine
reservation wages in standard search models (e.g., Mortensen, 1986).
Second, urban models predict that cities with more employers and work-
ers allow for better matches between the skill endowments of workers
and the skill requirements of jobs (Helsley and Strange, 1990). Third,
economic sociologists have stressed that social networks – which are of-
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

ten local – are important in finding new jobs. In line with this, proximity
to suitable jobs has been found to decrease joblessness, even within a
single city (Andersson et al., 2014) and displacement effects have been
found to be more severe in declining local economies (Jacobson et al.,
1993) and industries (Carrington, 1993).
Marshall’s original argument for what are now known as Marshal-
lian externalities was that a large local concentration of an industry in a
region reduces the risk of protracted unemployment for the specialized
subject to eA ≥ 0, eB ≥ 0 and 𝑒𝐴 + 𝑒𝐵 ≤ 1, where b represents the value of
leisure, r a discount rate, W(x) the NPV of accepting a wage offer of x
and then staying in this job indefinitely, and Fs (x) the likelihood of being
offered a wage of x or less. rV can be interpreted as the “rental income”
derived from the expected NPV of future search processes. Assuming
optimal search now and in the future, this equals the value a worker
( )
derives from leisure, b, minus the costs of search, 𝑐 𝑒𝐴 + 𝑒𝐵 , plus the
increase of expected NPV of future incomes due to search. 1

workers employed in that industry. This suggests that displacement will The reservation wage is the same in both sectors: 𝑉 = 𝑤∗𝐴 ∕𝑟 =
depend on how many local jobs exist that utilize a displaced worker’s 𝑤∗𝐵 ∕𝑟 = 𝑤∗ ∕𝑟.2 Given that a worker could enjoy leisure valued at b by
skills. Interestingly, although authors have studied the exit (Gathmann not searching at all, w∗ must be at least equal to b for the worker to par-
et al., 2014) and entry (Greenstone et al., 2010) of large economic es- ticipate in the labor market (i.e., search). The constrained maximization
tablishments to identify causal effects of Marshallian externalities by problem above now becomes:
exploiting the employment shocks these events create, this work has fo- [
cused on labor market pooling benefits to firms, not workers. Moreover, ( ) ∑ 𝜓𝑠 ( ) ∞ ( )
ma𝑥𝑒𝐴 ,𝑒𝐵 𝑏 − 𝑐 𝑒𝐴 + 𝑒𝐵 + 𝜎 𝑒𝑠 𝑥 − 𝑤 ∗ d 𝐹 𝑠 (𝑥 )
recent research in economic geography has studied how workers switch 𝑟 ∫𝑤∗
𝑠∈{𝐴,𝐵 }
industries and regions in the aftermath of shipyard closures in Denmark ]
(Holm et al., 2017), Germany and Sweden (Eriksson et al., 2016) but this ( )
− λ eA + eB − 1
has no implications regarding whether Marshallian externalities change
the effects of establishment closures.
To address this issue, we study how the local concentrations of the for w∗ ≥ b, eA ≥ 0, eB ≥ 0 and 𝜆 a Lagrangian multiplier. Optimal search
pre-displacement and related industries impact the careers of displaced is now determined by the following first-order conditions:3
workers. Do such concentrations affect the earnings drop after displace- ( ) ( ) { ∞ }
𝜕𝑐 𝑒∗𝐴 + 𝑒∗𝐵 𝜕𝜎 𝑒∗𝐴 𝜓𝐴 ( )
ment? Do they affect the length of unemployment spells? Do they change − ( ∗ ) + 𝑥 − 𝑤 ∗
d 𝐹 𝐴 ( 𝑥 ) − 𝜆 = 0, 𝑤∗ ≥ 𝑏
𝜕 𝑒 + 𝑒∗ 𝜕𝑒𝐴 𝑟 ∫𝑤∗
whether workers deal with displacement by switching industries or by 𝐴 𝐵
moving to other regions? And, do displaced workers respond to the Mar- ( ) ( ) { ∞ }
𝜕𝑐 𝑒∗ + 𝑒∗ 𝜕𝜎 𝑒∗𝐵 𝜓𝐵 ( )
shallian externalities offered by these local industry concentrations by − ( ∗𝐴 ∗𝐵) + 𝑥 − 𝑤 ∗ d 𝐹 𝐵 (𝑥 ) − 𝜆 = 0 , 𝑤 ∗ ≥ 𝑏
aligning their search efforts with these externalities? 𝜕 𝑒𝐴 + 𝑒𝐵 𝜕𝑒𝐵 𝑟 ∫𝑤∗

Optimal search thus equalizes the marginal returns to search in both

3. Model sectors. Consequently, at optimal effort levels, 𝑒∗𝐴 and 𝑒∗𝐵 , the following
must hold:4
To structure our empirical analyses, we build on a model of job ( ) ∞
𝜎 ′ 𝑒∗𝐴 𝜓𝐵 ∫𝑤∗ (𝑥 − 𝑤∗ )d𝐹𝐵 (𝑥)
search developed by Fallick (1992, 1993). In this model, unemployed ( ∗) = ∞ , 𝑤∗ ≥ 𝑏 (2)
workers divide their search efforts between two sectors. As in Fallick 𝜎 ′ 𝑒𝐵 𝜓𝐴 ∫𝑤∗ (𝑥 − 𝑤∗ ) d𝐹𝐴 (𝑥)
(1993), we will think of the first sector as the industry from which When the distribution of wage offers or arrival rates in sector A dete-
the worker was displaced and the second sector as consisting of other riorate compared to those in sector B, the right-hand side ratio increases.
suitable industries, i.e., industries that require similar skills as the pre- By assumption, 𝜎′ is positive and monotonically decreasing. Therefore,
displacement industry. We then proceed to give this model an explicitly to increase the left-hand side ratio, under optimal search, efforts will
( )
spatial dimension, by assuming that search efforts translate into – among 𝜎 ′ 𝑒∗𝐴
other things – a widening of the geographical scope of the search. shift from sector A to sector B. Moreover, given that ( ) is monoton-
𝜎 ′ 𝑒∗𝐵
Let there be two sectors s ∈ {A, B}, which are characterized by an
ically decreasing in 𝑒∗𝐴 , the ratio of sector A to sector B’s attractiveness
offer-arrival parameter 𝜓 s and a cumulative wage-offer distribution
has a one-to-one mapping to 𝑒∗𝐴 , guaranteeing that (2) has a unique so-
Fs (w). Search efforts, es , are sector-specific and increase the job-offer
(∑ ) lution.5
arrival-rate in a sector but are also costly, 𝐶 = 𝑐 𝑠 𝑒𝑠 . The arrival rate
of job offers is assumed to follow a Poisson distribution with an arrival
rate 𝛼 s that depends on the intrinsic, sector-specific offer-arrival param- 1
If a worker finds a job that pays a wage of x in each period, the present
eter 𝜓 s and the search intensity in sector s: value of this job offer equals 𝑊 (𝑥) = 𝑥∕𝑟, representing an increase of 𝑊 (𝑥) − 𝑉 =
( ) 𝑥∕𝑟 − 𝑉 over the NPV of engaging in search. Because the instantaneous offer
𝛼𝑠 = 𝜓𝑠 𝜎 𝑒𝑠 (1) arrival rate equals 𝜓 s 𝜎(es ), the expected net present value of searching with
∑ ( ){ ∞ }
effort levels es equals 𝑠 𝜓𝑠 𝜎 𝑒𝑠 ∫0 max[0, 𝑊 (𝑥) − 𝑉 ] d𝐹𝑠 (𝑥) .
The function 𝜎(es ) links offer arrival rates to search efforts. Each 2
This holds even in models where sectors have different layoff rates (Fallick,
worker has a total search budget of one unit of effort: Σs es ≤ 1. To re- 1992). The reason is that, at the reservation wage, workers are indifferent be-
ceive job offers, a non-zero effort is required. Beyond this initial effort, tween unemployed search and employment. Furthermore, if search costs are the
job offer arrival rates increase monotonically with effort but marginal same when employed or unemployed, workers can continue their search while
( ) ( )
returns are diminishing in each sector: 𝜎(0) = 0, 𝜎 ′ 𝑒𝑠 > 0, 𝜎 ′′ 𝑒𝑠 < 0. working. In this case, the reservation wage equals the value of leisure.
𝜕𝑐 (𝑒𝐴 +𝑒𝐵 ) 𝜕𝑐 (𝑒 +𝑒 ) 𝜕 (𝑒𝐴 +𝑒𝐵 ) 𝜕𝑐 (𝑒 +𝑒 ) 𝜕𝑐 (𝑒𝐴 +𝑒𝐵 )
While unemployed, workers maximize the net present value (NPV) 3
We use the fact that 𝜕𝑒𝐴
= 𝜕 𝑒 𝐴+𝑒 𝐵 = 𝜕 𝑒 𝐴+𝑒 𝐵 = .
( 𝐴 𝐵 ) 𝜕𝑒𝐴 ( 𝐴 𝐵) 𝜕𝑒𝐵
of searching for a job in the next period, V, by deciding how much search 𝜕𝑐 2 𝑒∗ +𝑒∗
( ) 𝜕 2 𝜎 𝑒∗
( ) 𝜓 { ∞ }
4
Denoting 𝑐 ′′ = 𝜕 𝑒∗ +𝑒∗ 𝐴𝜕 𝑒∗𝐵 +𝑒∗ , 𝜎𝑠′′ = 𝜕𝑒2 𝑠 and 𝑘𝑠 = 𝑟𝑠 ∫𝑤∗ (𝑥 − 𝑤∗ )d𝐹𝑠 (𝑥)
effort they want to dedicate to each sector and on a reservation wage, ( 𝐴 𝐵) ( 𝐴 𝐵) 𝑠

⎡0 1 1 ⎤
𝑤∗𝑠 , at which they will stop searching and accept a job. From standard
the bordered Hessian is given by ⎢1 −𝑐 ′′ + 𝑘𝐴 𝜎𝐴′′ −𝑐 ′′ ⎥, which – by
continuous-time wage-search theory (e.g., Mortensen, 1986), it follows ⎢ ⎥
⎣1 −𝑐 ′′ −𝑐 ′′ + 𝑘𝐵 𝜎𝐵′′ ⎦
that the worker maximizes the expected net income stream when: the assumption of decreasing marginal benefits to search efforts (𝜎𝑠′′ < 0) – en-
[ sures that the interior solution is a maximum, as long as marginal costs of search
( ) ∑ ( ) are increasing, constant or decreasing sufficiently slowly (i.e., 𝑐 ′′ > 𝑘𝐵 𝜎𝐵′′ ).
rV = ma𝑥𝑒𝐴 ,𝑒𝐵 𝑏 − 𝑐 𝑒𝐴 + 𝑒𝐵 + 𝜓𝑠 𝜎 𝑒𝑠 5
This follows from the fact that, given that 𝜎′′ < 0 and the first-order condi-
𝑠∈{𝐴,𝐵 }
{ }] tion related to the budget constraint implies that 𝑒∗𝐴 + 𝑒∗𝐵 = 1, the derivative of
∞
𝜎 ′ (𝑒∗ )
max [0, 𝑊 (𝑥) − 𝑉 ] d𝐹𝑠 (𝑥) log 𝜎 ′ 1−𝐴𝑒∗ with respect to 𝑒∗𝐴 is negative for any 0 ≤ 𝑒∗𝐴 ≤ 1.
∫0 ( 𝐴)

126
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Whenever a sector offers a job with a wage above the reservation

( )𝛿𝐴 ( )𝛿𝐵
wage, w∗ , search ends and workers exit unemployment through this sec- 𝜃𝐴 𝜃𝐵
𝐿 = ℎ𝛿𝐴 +𝛿𝐵 (1 − ℎ)1−𝛿𝐴 −𝛿𝐵
tor. Because the likelihood of such an event is independent of the time a 𝜃𝐴 + 𝜃𝐵 𝜃𝐴 + 𝜃𝐵
worker has spent searching, the destination-sector-specific hazard rate
is constant and equal to: Approximating ℎ = 1 − 𝑒−(𝜃𝐴 +𝜃𝐵 ) by 𝜃𝐴 + 𝜃𝐵 :
( ) [ ( )] ( ) 𝛿𝐴 ( )𝛿𝐵
𝜃𝑠 = 𝜎 𝑒∗𝑠 𝜓𝑠 1 − 𝐹𝑠 𝑤∗ , 𝑤∗ ≥ 𝑏 (3) ( )𝛿 +𝛿 ( )1−𝛿𝐴 −𝛿𝐵 𝜃𝐴 𝜃𝐵
𝐿 ≊ 𝜃𝐴 + 𝜃𝐵 𝐴 𝐵 1 − 𝜃 𝐴 − 𝜃𝐵
In principle, one could use a competing-risks model to approach this 𝜃𝐴 + 𝜃𝐵 𝜃𝐴 + 𝜃𝐵
problem empirically. However, we observe workers only once a year, ( )1−𝛿𝐴 −𝛿𝐵 𝛿𝐴 𝛿𝐵
for up to to three years after displacement. Consequently, our data on 𝐿 ≊ 1 − 𝜃𝐴 − 𝜃𝐵 𝜃𝐴 𝜃𝐵
survival are in discrete time. Standard continuous-time competing-risk
models are therefore less suited. Below, we adapt the derivations in If we choose a logistic function to relate hazard rates to observables,
𝑒𝑋𝛽𝑠
Jenkins (2005, pp. 103-105) to the context of the hazard rate in (3) to i.e. 𝜃𝑠 = , we obtain the likelihood function associated with
1+𝑒𝑋𝛽𝐴 +𝑒𝑋𝛽𝐵
show that the determinants of a worker’s hazard to exit unemployment a multinomial logit model:
through sector A or through sector B can be estimated approximately by ( )1−𝛿𝐴 −𝛿𝐵 ( )𝛿𝐴
a multinomial logit model. 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵 𝑒𝑋𝛽𝐴
𝐿≊ 1−
Let f(u, v) be the joint probability density function for the probability 1 + 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵 1 + 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵
that acceptable job offers arrive in sector A at time u and in sector B at ( )𝛿𝐵
𝑒𝑋𝛽𝐵
time v. The hazard of exiting unemployment through sector A, i.e., the 1 + 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵
probability that a worker will have accepted a job in sector A by the end
of a one time-period, is given by:
( )1−𝛿𝐴 −𝛿𝐵 ( )𝛿𝐴
𝑒𝑋𝛽𝐴
1 ∞
1
𝑃 (𝑢 < min(𝑣, 1)) = 𝑓 (𝑢, 𝑣)d𝑣 d𝑢 (4) 𝐿≊
∫0 ∫𝑢 1 + 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵 1 + 𝑒𝑋𝛽𝐴 + 𝑒𝑋𝛽𝐵
( )𝛿𝐵
As common in competing risks models, we assume that, conditional 𝑒𝑋𝛽𝐵
on observables, the destination specific continuous hazard rate functions (7)
1 + 𝑒 𝐴 + 𝑒𝑋𝛽𝐵
𝑋𝛽
are independent. Eq. (4) can then be rewritten as:
{ }
1 1 ∞ The geography of search
𝑃 (𝑢 < min(𝑣, 1)) = 𝑓𝐴 (𝑢)𝑓𝐵 (𝑣) d𝜈 + 𝑓𝐴 (𝑢)𝑓𝐵 (𝑣) d𝜈 d𝑢 (5)
∫0 ∫𝑢 ∫1
In order to add a spatial dimension to the search process, we assume
Let hs be a discrete hazard rate for exiting unemployment through that the sector-specific intrinsic offer rates, 𝜓 s , or the wage-offer dis-
sector s, i.e., the likelihood that an acceptable job offer arrives in sector tributions, Fs (w), or both, depend on the local labor market conditions
s before the end of the period. The second part of Eq. (5) now simplifies in sector s. In particular, holding local conditions for sector B constant,
to: more favorable conditions for sector A will directly and positively effect
1 ∞ ( ) 1 ( ) 𝜃 A , but not 𝜃 B . However, the improvement in local conditions in sector
𝑓𝐴 (𝑢)𝑓𝐵 (𝑣) d𝜈 d𝑢 = 1 − ℎ𝐵 𝑓 (𝑢 ) d 𝑢 = ℎ 𝐴 1 − ℎ 𝐵
∫0 ∫1 ∫0 𝐴 A will induce workers to reallocate search efforts from sector B to sec-
( )
= ℎ𝐴 1 − ℎ𝐵 tor A. Consequently, local conditions in A will still indirectly affect 𝜃 B ,
namely, through the way workers divide search efforts. Eq. (2) shows
Let Ss (x) be the survival function for sector s, i.e., the likelihood that that this effect will be negative: the better the local conditions in sector
no acceptable offer has arrived from sector s until time x. Given that the A are, the less a worker will search in sector B. This, in turn, decreases
hazard functions are constant over time, the first part of Eq. (5) can now 𝜃 B , the hazard of exiting unemployment through sector B.
be written as:6 How would these search efforts be reflected in observable charac-
1 1 𝜃𝐴 ( ) teristics of workers’ careers? By increasing search efforts, workers can
𝑓 (𝑢)𝑓𝐵 (𝑣) d𝜈 d𝑢 = ℎ − 1 − ℎ𝐵 ℎ𝐴 (6)
∫0 ∫𝑢 𝐴 𝜃𝐴 + 𝜃𝐵 sample jobs from a wider sets of firms. We propose that one of the ways
where h represents the likelihood that the worker finds a job in either in which this manifests itself is through an increase in the geograph-
of the two sectors before the end of the period and 𝜃 s the instantaneous ical scope of search. For instance, workers could spend extra time to
hazard of finding a job in sector s. Furthermore, we have used that with attend job interviews outside the region. Similarly, social networks may
only one time period, the discrete hazard rate is the complement of help identify job opportunities outside the region. However, given that
the survival function at the end of that period: ℎ = 1 − 𝑆𝐴 (1)𝑆𝐵 (1) = 1 − these networks are often local, this may involve reaching out to friends
𝑆 (1), where S(.) represents the joint survival function for the hazards of and acquaintances that are somewhat removed in one’s social network.
finding a job in A or B. Putting both pieces together, Eq. (5) becomes: As a consequence, identifying nonlocal, as opposed to local jobs, will
{ } typically require more effort. Note that we assume that workers pre-
1 1 ( ) 𝜃𝐴 ( )
𝑓 𝐴 (𝑢 ) 𝑆𝐵 (𝑣)𝜃𝐵 d𝑣 d𝑢 = ℎ𝐴 1 − ℎ𝐵 + ℎ − 1 − ℎ𝐵 ℎ𝐴 dominantly direct their search efforts at sectors, not regions. In Online
∫0 ∫𝑢 𝜃𝐴 + 𝜃𝐵 Appendix B, we provide some empirical results that show that this is a
𝜃𝐴 plausible assumption.
= ℎ We incorporate this reasoning into the model by modifying Eq. (1) to
𝜃𝐴 + 𝜃𝐵
add a geographical dimension to the arrival rates of suitable wage of-
The probability that the worker receives an acceptable offer from
fers. In particular, let offers from sector s originate from outside the
sector B first is analogous. Finally, the probability of receiving no ac-
worker’s home region with probability 𝜌(es |Xs ). The hazard of exiting
ceptable offer at all before the end of the period is simply 1 − ℎ. Conse-
unemployment through sector s in her home region, 𝜃 0s , now becomes:
quently, the likelihood of observing 𝛿 A individuals accepting job offers
in sector A and 𝛿 B individuals accepting offers in sector B is: ( )( ( ))[ ( )]
( )𝛿𝐴 ( )𝛿𝐵 𝜃0𝑠 = 𝜓𝑠 𝜎 𝑒𝑠 1 − 𝐹𝑠 𝑤∗ 1 − 𝜌 𝑒𝑠 |𝑋𝑠 (8)
𝜃𝐴 𝜃𝐵
𝐿 = (1 − ℎ)1−𝛿𝐴 −𝛿𝐵 ℎ ℎ
𝜃𝐴 + 𝜃𝐵 𝜃𝐴 + 𝜃𝐵 and the hazard of exiting through jobs in sector s outside the home re-
gion, 𝜃 1s , equals:
( )( ( )) ( )
6
See Online Appendix A for a full derivation. 𝜃1𝑠 = 𝜓𝑠 𝜎 𝑒𝑠 1 − 𝐹𝑠 𝑤∗ 𝜌 𝑒𝑠 |𝑋𝑠 (9)

127
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

𝜌 thus maps search efforts onto the interval (0, 1). We will assume The derivative of this expression with respect to local conditions in
that 𝜌 decreases monotonically in Xs , a vector that captures how favor- sector B is given by:
𝜕𝜌 ( ( ) )
able local conditions are in sector s. That is, we will assume that 𝜕𝑋 < 0,
𝑠 𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴 ( )
such that favorable local conditions raise the likelihood that acceptable 𝜕 log ( ( ) ) ∕𝜕 𝑋𝐵 ≂ 𝜕 𝑋𝐴 𝛽𝐴1𝐴 + 𝑋𝐵 𝛽𝐵1𝐴 ∕𝜕 𝑋𝐵 ,
1 − 𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴
offers will arrive from within the region as opposed to from outside the
region. Moreover 𝜌 is assumed to increase in es to reflect the greater which evaluates to:
efforts that job offers from outside the region require.7
1
In the empirical analyses, we will equate one of the two sectors in ( ( ) )( ( ( ) ))
𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴 1 − 𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴
the model with the 5-digit industry from which workers are displaced. ( ( ) ) ( )
Henceforth, we will refer to this industry as the “pre-displacement in- 𝜕𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴 𝜕𝑒𝐴 𝑋𝐴 , 𝑋𝐵
( ) ≂ 𝛽𝐵1𝐴
dustry” or a worker’s “old industry.” The other sector consists of other 𝜕𝑒𝐴 𝑋𝐴 , 𝑋𝐵 𝜕𝑋𝐵
industries that provide suitable jobs, namely those that are related to
the pre-displacement industry. The upshot of Eqs. (8) and (9) is that we Given that the first ratio is always positive, and the second is pos-
can infer how workers allocate search efforts between these two sec- itive by assumption, a significant and negative estimate for 𝛽𝐵1𝐴 implies
tors from workers’ geographic mobility. In particular, the model has the 𝜕𝑒𝐴 (𝑋𝐴 ,𝑋𝐵 )
𝜕𝑋
< 0. Therefore, if we find that favorable local conditions in
𝐵
following testable predictions:
sector B reduce the log-odds of exiting through nonlocal instead of local
1. Favorable local conditions in the pre-displacement industry (in re- jobs in sector A (i.e., 𝛽̂𝐵1𝐴 < 0), we can infer that these favorable local
lated industries) will increase the likelihood of finding jobs in this conditions must have negatively impacted the efforts workers dedicated
industry (in these industries). to search in sector A.
2. Conditional on the local conditions in related industries (in pre- Hypotheses 3 can be derived from an inspection of the log-odds for
displacement industries), favorable local conditions in the pre- exiting unemployment through nonlocal jobs in sector A instead of stay-
displacement industry (in related industries) will decrease the rela- ing unemployed:
tive risk of finding nonlocal jobs compared to local jobs outside the 𝜃1 𝐴
log ≂ 𝑋𝐴 𝛽̃𝐴1𝐴 + 𝑋𝐵 𝛽̃𝐵1𝐴 ,
pre-displacement industry (in the pre-displacement industry). 1 − 𝜃0 𝐴 − 𝜃1 𝐴 − 𝜃0 𝐵 − 𝜃1 𝐵
3. Conditional on the local conditions in related industries (in pre-
displacement industries), favorable local conditions in the pre- where coefficients 𝛽̃𝐴1𝐴 and 𝛽̃𝐵1𝐴 have been normalized against the coeffi-
displacement industry (in related industries) will decrease the likeli- cients of the hazard to remain unemployed. Using Eqs. (8) and (9), the
hood of finding nonlocal jobs outside the pre-displacement industry derivative of this expression with respect to local conditions in sector B
(in the pre-displacement industry) compared to staying unemployed. can be written as:
( )
𝜕𝑒𝐴 𝑋𝐴, 𝑋𝐵
Prediction 1 derives from the fact that the quality of local job-offers Ψ1
and/or arrival rates increase in a sector as a direct effect of better local 𝜕𝑋𝐵
( ) ( ) ( )
conditions in that sector. This effect is augmented by the fact that better 𝜕𝑒𝐴 𝑋𝐴, 𝑋𝐵 𝜕𝑒𝐴 𝑋𝐴, 𝑋𝐵 𝜕𝑘𝐵 𝑋𝐵 ( ( ))
Ψ2 − Ψ3 + 𝜎 𝑒𝐵 𝑋𝐴, 𝑋𝐵
local conditions will also spur greater efforts to search in the sector, 𝜕𝑋𝐵 𝜕𝑋𝐵 𝜕𝑋𝐵
which raises the likelihood of receiving acceptable job offers. The effect + ≂ 𝛽̃𝐵1𝐴
Ψ4
on whether acceptable offers will be local or nonlocal is ambiguous,
because 𝜌 decreases due to better local conditions, but increases because where:
( ) ( )
of greater efforts. However, local conditions in sector A should neither 𝑘𝑠 𝑋𝑠 = 𝜓𝑠 1 − 𝐹𝑠 (𝑤∗ )
𝜕𝜎 (𝑒𝐴 (𝑋𝐴, 𝑋𝐵 )) 𝜕𝜌(𝑒𝐴 (𝑋𝐴, 𝑋𝐵 ),𝑋𝐵 )
directly affect the ratio of nonlocal to local job-finding hazards in sector Ψ1 = 𝜎 𝑒 𝑋1 𝑋 + 𝜌 𝑒 𝑋1 𝑋
( 𝐴 ( 𝐴, 𝐵 )) 𝜕𝑒𝐴 (𝑋𝐴, 𝑋𝐵 ) ( 𝐴 ( 𝐴, 𝐵 )) 𝜕𝑒𝐴 (𝑋𝐴, 𝑋𝐵 )
B nor of the ratio of finding a job in sector B to remaining unemployed. ( ) 𝜕𝜎 (𝑒 (𝑋 𝑋 ))
Such cross-over effects nevertheless arise, because favorable conditions Ψ2 = 𝑘𝐴 𝑋𝐴 𝜕𝑒 𝐴 𝑋 𝐴,𝑋 𝐵
𝐴 ( 𝐴, 𝐵 )
in sector A will draw search efforts from sector B to sector A as implied ( ) 𝜕𝜎 (1−𝑒 (𝑋 𝑋 )) 𝜕 (1−𝑒𝐴 (𝑋𝐴, 𝑋𝐵 )) ( ) 𝜕𝜎 (𝑒 (𝑋 𝑋 ))
Ψ3 = −𝑘𝐵 𝑋𝐵 𝜕 1−𝑒 𝐴 𝑋 𝐴,𝑋 𝐵 = 𝑘𝐵 𝑋𝐵 𝜕𝑒 𝐵 𝑋 𝐴,𝑋 𝐵
by Eq. (2). ( )( ( 𝐴 (( 𝐴, 𝐵 )) )) 𝜕𝑒𝐴 (𝑋𝐴, ( 𝑋𝐵 )) ( ( )) 𝐵 ( 𝐴, 𝐵 )
To see this, consider that the model specification in (7) implies the Ψ4 = 1 − 𝑘𝐴 𝑋𝐴 𝜎 𝑒𝐴 𝑋𝐴, 𝑋𝐵 − 𝑘𝐵 𝑋𝐵 𝜎 𝑒𝐵 𝑋𝐴, 𝑋𝐵
following log-odds for exiting unemployment through nonlocal instead If search efforts were not responsive to local conditions, i.e.,
of local jobs in sector A: 𝜕𝑒𝐴 (𝑋𝐴, 𝑋𝐵 ) 𝜕𝑘 (𝑋 )
𝜕𝑋𝐵
= 0, 𝛽̃1𝐴 should have the same sign as 𝐵 𝐵 and therefore
𝐵 𝜕𝑋𝐵
𝜃1 𝐴 be positive. A negative effect of conditions in sector B on the relative
log ≂ 𝑋𝐴 𝛽𝐴1𝐴 + 𝑋𝐵 𝛽𝐵1𝐴 , (10)
𝜃0 𝐴 risk of finding nonlocal jobs in sector A vis-à-vis remaining unemployed
𝜕𝑒 (𝑋 𝑋 )
where Xs quantifies the quality of local conditions in sector s and param- would therefore imply that 𝐴 𝜕𝑋𝐴, 𝐵 < 0.8 Note, that a similar observa-
𝐵
eters are subscripted by this sector and superscripted by the hazard rate tion can be made for the effects on the relative risk of finding local jobs
to which they refer (0 codes exits through local, 1 through nonlocal jobs in sector A instead of remaining unemployed. In this case, the second
and A and B code the sector in which the job was found). Furthermore,
the coefficients of the reference category (local exits through sector A), ( ) 𝜕𝑒 (𝑋 𝑋 )
𝛽𝐴0𝐴 and 𝛽𝐵0𝐴 , are normalized to zero and ≂ indicates an equality by as-
8
In principle, this finding could also imply that Ψ3 − Ψ2 𝐴 𝜕𝑋𝐴, 𝐵 >
𝐵
𝜕𝑘𝐵 (𝑋𝐵 ) ( ( ))
sumption of the logistic functional form. Using (8) and (9), (10) implies: 𝜎 𝑒𝐵 𝑋𝐴, 𝑋𝐵 , requiring that the indirect effect of the shift in search
𝜕𝑋
( ( ) ) 𝐵
efforts due to better local conditions in sector B on total job finding rates is neg-
𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴
log ( ( ) ) ≂ 𝑋𝐴 𝛽𝐴1𝐴 + 𝑋𝐵 𝛽𝐵1𝐴 . ative and in excess of the positive direct effect on job-offers. This would mean
1 − 𝜌 𝑒𝐴 𝑋𝐴 , 𝑋𝐵 , 𝑋𝐴 that the derivative of the entire denominator of the log-odds were negative. That
is, better conditions in a sector would have to lead to an increase in the likeli-
hood that workers remain unemployed. This is not impossible: if the sector in
7
Note that we do not specify whether efforts and favorable local conditions which search conditions improve offers much higher wages but much worse job
increase job offer arrival rates or lead to better wage-offer distributions. Because, arrival rates, workers might be tempted to shift search efforts to this sector be-
without loss of generality, we can think of wages net of commuting and/or cause of higher expected wages (but lower employment chances) in a way that
relocation costs, the optimization problem of the worker remains unaffected by outdoes the improved conditions. However, we believe such an outcome to be
the fact that some jobs are located outside the region. unlikely in general.

128
F.M.H. Neffke et al.
𝜕𝜌(𝑒𝐴 (𝑋𝐴, 𝑋𝐵 ),𝑋𝐵 )
term in Ψ1 is replaced by − 1−𝜌 𝑒 1𝑋 𝑋 , dampening
( 𝐴 ( 𝐴, 𝐵 )) 𝜕𝑒𝐴 (𝑋𝐴, 𝑋𝐵 )
the negative effect of effort reallocation. Intuitively, this happens, be-
cause decreasing search efforts will reduce the spatial scope of search in
sector A, which, in turn, will increase local job-finding rates. We will test
whether the cross-dependencies predicted in hypotheses 2 and 3 exist
in reality at the end of section 6.
4. Data
Our data are taken from the German Employee History database
(Beschäftigtenhistorik, EH, see Bender et al. (2000) for details). The EH
database is based on Germany’s social security records. Our version of
these data provides yearly information on an individual’s daily wage,
deflated to 2005 EUR, occupation, work status (i.e., full-time employed,
part-time employed, or in apprenticeship), gender, and age. The EH also
contains anonymized identifiers that allow us to follow individuals over
time. Moreover, the EH contains information about the industry and lo-
cation of each establishment. Because of changes in the industry classifi-
cation system, we limit our analyses to the years 1999 to 2008. Further-
more, we focus on male, full-time employees between the ages of 25 and
50 and drop apprentices. To ensure a uniform definition of success in
post-displacement job-search across workers, we ignore part-time jobs.
This strategy may lead to overstated wage losses for displaced workers.
We therefore rerun our analyses while including all post-displacement
part-time jobs. Apart from some reductions in the effect of displacement
on wages and nonemployment rates, these estimations yield virtually
the same results as those presented hereafter (see Online Appendix D).
A drawback of social security records is that they do not cover in-
dividuals who are exempt from social security contributions, such as
civil servants, soldiers, self-employed workers, entrepreneurs and un-
paid family workers. In total, these workers constitute about 20 percent
of the German labor force (Herberger and Becker, 1983). When we use
the term “employed”, we therefore refer to people employed in jobs
with social security coverage. Similarly, although the main reason indi-
viduals drop out of the data is because they have become unemployed
or inactive, some may also have returned to school, received civil ser-
vant status, started their own businesses, and so on. We therefore use
the term “nonemployment” instead of unemployment to refer to work-
ers who leave jobs with social security coverage. Online Appendix F
addresses some issues arising from this definition in further detail.
As displaced workers, we select all workers who have lost their jobs
in establishment closures. Closures are identified with the help of a vari-
able created by Hethey and Schmieder (2010). These authors marked
each disappearance of an establishment identifier from the EH as a
potential closure event. However, when analyzing the labor outflows
from these establishments, they found that only about 40% of estab-
lishments of four employees or larger with a disappearing identifier can
be regarded unambiguously as closures rather than mere administra-
tive changes in establishment identifiers. In the remaining 60%, large
shares of the disappearing establishment’s workforce move to the same
new employer, which suggests that some kind of corporate connection
(such as take-overs or identifier recodings) exists between the old and
the new establishment. In what follows, we will consider establishment-
identifier disappearances to signal closure events, if the establishment
had at least 10 employees in the year before the closure, and if fewer
than 30% of its workers move to the same other establishment in the
year after the closure (see Online Appendix F for a discussion of alter-
native definitions of closure events). We then gather all workers who left
one of these establishments during the year they closed down. Of these
workers, we select those who, prior to the displacement event, (a) had
at least six years of work experience, (b) three years of industry expe-
rience and (c) one year of establishment tenure. These three conditions
ensure that workers have had enough time to find well-matching jobs
and gain relevant work experience, such that their industry affiliation
is a good reflection of their (industry-specific) skills. Moreover, insist-
129
Journal of Urban Economics 108 (2018) 124–140
ing on over one year of establishment tenure avoids selecting workers
who were hired for reasons directly related to the closure. We then fol-
low these workers for the period starting six years before and ending
three years after the closure. These conditions limit us to establishment
closures between 2003 and 2005.
The empirical analyses are divided into two parts. First, we will esti-
mate the causal effects of job displacement on workers’ career trajecto-
ries and to what extent the local industry mix moderates these effects.
In particular, we will quantify the effect of displacement events, not just
on employment rates and wages, but also on workers’ geographical and
industrial mobility. This will help determine whether there is evidence
for the existence of Marshallian benefits in job search. We describe these
analyses in section 5.
Second, we will explore whether displaced workers internalize these
Marshallian benefits in their search strategies. In particular, we will test
the hypotheses formulated in section 3 using a multinomial logit estima-
tion. The core of these hypotheses is that, all else equal, workers should
reduce search efforts in one sector when the Marshallian externalities
in the other sector increase. These reductions in search efforts should
manifest themselves in a contraction of the spatial scope of search. The
second part of the empirics – summarized in section 6 – tests this pre-
diction.
5. Displacement effects
Related industries
In the model of section 3, workers divide search efforts between two
sectors: the pre-displacement industry and a second sector consisting of
industries that are closely related to the pre-displacement sector in terms
of their skill requirements. To define the set of related industries that
constitute this second sector, we use the skill-relatedness index proposed
by Neffke et al. (2017). This index is calculated as the observed labor
flows between two industries, divided by the labor flows that would
have been expected had workers switched industries randomly. That is,
let Fij be the number of workers who change jobs from establishments
in industry i to establishments in industry j. The relatedness between i
and j is now defined as:
𝐹𝑖𝑗 ∑∑
𝑅𝑖𝑗 = ∑ ∑ 𝐹𝑘′ 𝑙 ′ (11)
𝑘≠𝑗 𝐹𝑘𝑗 𝑙≠𝑖 𝐹𝑖𝑙 𝑘′ 𝑙′ ≠𝑘′
and Rii ≡ 0: industries are by definition not skill-related to themselves.
To enhance the precision of the index, we construct these labor flows
using information on all full-time employed men and women between
the age of 18 and 65. Similar inter-industry relatedness indices have
been used in a variety of studies (Greenstone et al., 2010; Dauth, 2010;
Baptista and Costa, 2012; Neffke and Henning, 2013; Timmermans and
Boschma, 2013; Csáfordi et al., 2018)
Because inter-industry labor-flow connections are extremely sparse
– in any given year, over 80% of industry pairs display no labor flows
at all – this method provides clearly delineated labor markets. To avoid
mechanical relations between the R-matrix and the careers of displaced
workers, we remove from Fij all workers who at any point between 1999
and 2008 had been employed in an establishment that closes down in
this same period.9 Finally, we calculate this R-matrix for each year be-
tween 1999 and 2008, take its average across years, and symmetrize the
9 Neffke et al. (2017) show that these matrices are all but invariant across
years and highly similar across a number of broad occupational and wage group-
ings. Moreover, flows of workers who change jobs over long or short distances
yield all but indistinguishable R-matrices. This suggests that the patterns ex-
pressed in these matrices express some fundamental, non-idiosyncratic similar-
ities in job tasks across industries.
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

resulting matrix by averaging its elements with those of its transpose.10 Consequently, the two sectors we distinguish (the old industry and re-
We refer to this averaged and symmetrized matrix as 𝑅̄ . lated industries) represent similarly important reservoirs of new jobs.
Variations of this threshold and definitions analogous to Eqs. (12) and
Local conditions (13) based on employment levels, instead of shares, yield similar results
(available on request).
Our main interest is the role Marshallian externalities play in the
post-displacement careers of workers who lose their job in establish- Estimation strategy
ment closures. Therefore, we define the local conditions in the model of
section 3 in terms of local industrial concentration patterns. In particu- Most job separations occur when workers decide it is time to pur-
lar, we use the local employment shares of the pre-displacement and of sue career opportunities elsewhere, or when their employers make this
related industries to categorize industry-region combinations into dif- decision in their stead. As a consequence, job separations are often en-
ferent classes. dogenous to the expectations about a workers’ career prospects at their
As regional units, we use Germany’s 141 labor market areas as de- current firm. An exception is job separations due to establishment clo-
fined by Kosfeld and Werner (2012). We start by dividing locations into sures. Such separations are typically unrelated to the performance and
three types, reflecting regions where the worker’s old (O) industry repre- career aspirations of individual workers and have, therefore, been con-
sents a small, moderate or large share of regional employment. To do so, sidered to be exogenous from a worker’s perspective (e.g., Gibbons and
we define the following dummy group for a worker who got displaced Katz, 1991; Jacobson et al., 1993; Couch and Placzek, 2010; Schwerdt,
from industry i in region r and year t: 2011). Using a sample of displaced workers should thus mitigate con-
( ) cerns about workers self-selecting into career changes as long as dis-
𝐸
𝐿
𝑂𝑖𝑟𝑡 = 𝐼 ∑ 𝑖𝑟𝑡 ≤ 𝜁1 placement is uncorrelated with worker characteristics.
𝑗 𝐸𝑗𝑟𝑡 To enhance the plausibility of the exogeneity assumption, we com-
( )
𝐸 pare displaced workers to observationally similar non-displaced work-
𝑀
𝑂𝑖𝑟𝑡 = 𝐼 𝜁1 < ∑ 𝑖𝑟𝑡 ≤ 𝜁2 ers, using a combination of propensity-score matching and regression
𝑗 𝐸𝑗𝑟𝑡
( ) analysis. To be precise, we follow Ho et al. (2007) and use matching
𝐸 as a pre-screening method to reduce the dependence of the treatment
𝐻
𝑂𝑖𝑟𝑡 = 𝐼 ∑ 𝑖𝑟𝑡 > 𝜁2 (12) variable (in our case, displacement) on worker characteristics. Such
𝑗 𝐸𝑗𝑟𝑡
pre-screening has several advantages. Firstly, because the procedure is
𝐸 based on only pre-displacement covariates, it does not introduce selec-
where ∑ 𝑖𝑟𝑡 is the regional employment share of the worker’s old in-
𝑗 𝐸𝑗𝑟𝑡 tion biases. Secondly, by ensuring a common support of treated and
dustry in year t (not counting the employment in the establishments
untreated individuals, pre-screening avoids inference that is based on
that close down). Furthermore I(.) is an indicator function that evalu-
inter- or extrapolation to parts of the covariate space where we do not
ates to 1 if its argument is true. Finally, 𝜁 1 and 𝜁 2 are chosen such that
observe any displaced (or nondisplaced) workers. Thirdly, because the
all categories represent an equal number of observations in our sample.
pre-screening ensures that displacement is orthogonal to the exogenous
Analogously, we group region-industry cells by the local employ-
covariates, we don’t need to make any parametric assumptions about
ment share of industries related to the pre-displacement industry (“Al-
how such covariates enter the data-generating process.11 As a conse-
ternative” industries):
quence, pre-screening mitigates misspecification issues related to the
( )
𝑟𝑒𝑙
𝐸𝑖𝑟𝑡 exact functional form through which these covariates are modeled in
𝐿 ′
𝐴𝑖𝑟𝑡 = 𝐼 ∑ ≤ 𝜁1 the regression equation (Ho et al., 2007). The cost of pre-screening the
𝑗 𝐸𝑗𝑟𝑡
( ) data is that the estimated effects represent average effects for the subset
𝑟𝑒𝑙
𝐸𝑖𝑟𝑡
𝑀 ′ ′ of displaced workers instead of for the population as a whole.
𝐴𝑖𝑟𝑡 = 𝐼 𝜁1 < ∑ ≤ 𝜁2
𝑗 𝐸𝑗𝑟𝑡
( ) Matching
𝑟𝑒𝑙
𝐸𝑖𝑟𝑡
𝐴𝐻
′
𝑖𝑟𝑡 = 𝐼 ∑ > 𝜁2 (13)
𝐸
𝑗 𝑗𝑟𝑡 Our matching strategy closely follows the one in Nedelkoska et al.
′ ′ (2015).12 For each displaced worker who meets the criteria listed in
𝜁1 and 𝜁2 once again divide workers into equally sized groups and Section 4, we try to find a statistical twin among the non-displaced
𝑟𝑒𝑙 represents the employment in region r and year t in industries
𝐸𝑖𝑟𝑡 workers by means of propensity-score matching. Statistical twins are
closely related to industry i, where “closely related” refers to industries drawn from a donor pool that observes the same pre-displacement re-
for which the skill-relatedness to the worker’s old industry i exceeds a strictions as the ones imposed on displaced workers, with the additional
threshold, 𝜉. That is: requirement that they do not experience any displacement events in the
∑ ( )
𝑟𝑒𝑙
𝐸𝑖𝑟𝑡 = 𝐸𝑘𝑟𝑡 𝐼 𝑅̄ 𝑖𝑘 > 𝜉 (14) 1999-2008 period.13 We estimate workers’ propensity to experience a
𝑘≠𝑖

Employment in Eqs. (13) and (14) is again measured in the displacement 11

year, excluding employment in establishments that close down. We use
a threshold value of 𝜉 = 3, which implies that the observed labor flows
between an industry and the pre-displacement industry are at least three
times as large as the random benchmark. The reason for this choice is
that, at this threshold, related industries absorb about the same share of
displaced workers (29%) as the pre-displacement industry itself (27%).
10
To be precise, we first use the following transformation to reduce skew:
𝑅∗ = 𝑅𝑅+1 , which maps the values of R from the interval [0, ∞) onto the interval
[0,1). This ensures that the averages are not overly affected by right-tail outliers.
The threshold value for R of 3 we use in this paper corresponds to a threshold
of 3/4 for R∗ .
For instance, mobility decisions will depend on a worker’s age. However,
because the functional relation between mobility and age may be complex, it
is hard to correct for this by simply controlling for worker age. By matching
displaced to nondisplaced workers, we select a sample of workers in which dis-
placement is orthogonal to age. Consequently, in this sample, a worker’s age
cannot confound the estimated displacement effect, irrespective of the exact
functional form through which mobility depends on age.
12
Nedelkoska et al. (2015) study occupational mobility of displaced workers
and the extent to which the need for skill-adjustments amplifies the effect of
displacement.
13
Note that this means that we do not impose any further restrictions on the
post-displacement careers of the donor pool. As a consequence, most statisti-
cal twins will not change employer in the year of displacement. An alternative
would be to match displaced workers to nondisplaced job-separators. However,

130
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table 1
Balance of matched sample.

selected population matched sample

treated control % bias treated control % bias

share rel. emp. 4.04% 4.62% -13.9 4.08% 4.07% 0.3

share old ind. emp. 0.72% 1.54% -28.8 0.72% 0.69% 3.3
age 39.8 39.7 1.7 39.8 39.8 0.8
edu (ND) 10.05% 10.47% -1.4 10.20% 10.35% -0.5
edu (VT) 63.93% 65.20% -2.7 65.35% 65.81% -1.0
edu (HS) 0.52% 0.65% -1.7 0.52% 0.46% 0.8
edu (HS+VT) 2.46% 4.01% -8.7 2.47% 2.43% 0.3
edu (C) 2.96% 4.60% -8.6 3.03% 3.02% 0.1
edu (U) 3.09% 6.39% -15.6 3.16% 3.21% -0.3
edu (miss.) 17.00% 8.69% 25.0 15.27% 14.72% 1.5
log(reg. size) 14.8 15.7 -14.7 15.2 15.2 -0.3
industry experience 9.2 10.6 -23.0 9.7 9.6 1.2
regional experience 12.4 13.7 -19.5 12.9 12.9 -0.5
establishment tenure 7.9 9.6 -28.6 8.3 8.3 0.0
year: 2005 38.86% 33.65% 10.8 38.50% 38.50% 0.0
year: 2006 35.06% 33.42% 3.5 35.13% 35.13% 0.0
year: 2007 26.09% 32.93% -15.0 26.38% 26.38% 0.0
wage (4 yrs pre-D.) 84.2 100.7 -32.1 88.7 88.4 0.5
wage (3 yrs pre-D.) 86.4 105.0 -34.7 90.0 89.6 0.8
wage (2 yrs pre-D.) 89.5 109.1 -35.2 94.0 94.1 -0.2
wage (1 yr pre-D.) 90.7 111.9 -36.0 93.0 95.5 -4.9
wage (at D.) 91.9 114.3 -36.5 93.9 97.4 -6.5
wage (1 yr post-D.) 50.8 110.4 -87.8 52.7 92.9 -66.7
wage (2 yrs post-D.) 60.5 108.5 -70.5 62.6 90.8 -46.6
wage (3 yrs post-D.) 63.0 106.6 -64.0 65.1 89.4 -40.0

The selected population refers to all individuals that meet the criteria outlined in
Section 4: full-time employees with at least (1) six years of work experience, (2) three
years of industry experience and (3) one year of establishment tenure. For matched
nondisplaced workers, we also require that they are not displaced at any time in the
1999-2008 period. “Share rel. emp.” refers to the share of skill-related employment
in the region at the time of (virtual) displacement as defined in Eq. (14). “Share old
ind. emp.” refers to the regional employment share of the pre-displacement industry.
Wages are real wages, denominated in 2005 EUR, at the specified number of years
before or after the displacement event (D.). Age, experience and tenure are measured
in years.

displacement event with a probit model that uses as explanatory vari-
ables a worker’s education, age, years of general, industry, and regional
work experience, as well as establishment tenure. To avoid paramet-
ric assumptions, age and experience variables enter as dummy groups.
Furthermore, we control for regional economic conditions by adding
the regional employment shares and squared values thereof in the pre-
displacement and in related industries in the year before the establish-
ment closes down. Most importantly, however, we use lags 6 to 2 of
pre-displacement wages and the logarithm of wage growth between 5
and 2 years before the displacement event to capture a worker’s pre-
displacement wage curve. Because this curve reflects rewards for both
observed and unobserved worker characteristics, it helps control for un-
observed characteristics that might affect post-displacement wage dy-
namics and mobility decisions. Matching workers with similar pre-event
wage curves, therefore, allows us to establish plausible counterfactual
careers for displaced workers. Finally, we match exactly on establish-
ment tenure and displacement year. After using nearest-neighbor match-
ing and dropping all observations that are outside the matching’s com-
mon support, we are left with a sample of 45,344 worker pairs.
Table 1 compares the means of the matching variables and wage
paths between displaced and non-displaced workers in the overall popu-
lation and in the selected sample. Individual characteristics of displaced
this design would not estimate the treatment effect of job displacement, i.e., the
difference between a displaced worker’s observed career path and her counter-
factual career path, had she not been displaced. Instead, we would end up es-
timating the difference between two treatments: the (exogenous) displacement
event on the one hand and (a most likely endogenous) regular job separation on
the other hand.
and non-displaced workers are much more closely aligned in the sam-
ple than in the population as a whole. For all pre-displacement vari-
ables, differences in means between displaced and non-displaced are
well below 5%. Note that pre-displacement wages are particularly well-
balanced, with biases below 1%.14 In as far as prior wages reflect a
worker’s productivity, the strong balance on these variables suggests
that there is little cause for concern that unobserved worker quality will
bias our results.
Findings
To assess the overall effects of displacement on earnings, wages, non-
employment and mobility decisions, we follow Schwerdt (2011) and
combine matching with the difference-in-differences framework intro-
duced to the displacement literature by Jacobson et al. (1993). That is,
we estimate the following equation:
∑
3 ∑
3
𝑦𝑚𝑡 = 𝜏1𝑘 𝑇𝑚𝑡
𝑘
+ 𝜏2𝑘 𝑇𝑚𝑡
𝑘
𝐷𝑚𝑡 + 𝑋𝑚𝑡 𝛽 + 𝛼𝑚 + 𝛿𝑡∗ (𝑚) + 𝜖𝑚𝑡 (15)
𝑘=−3 𝑘=−3
where the subscript t refers to the (calendar) year and t∗ (m) to the year
of the establishment closure. Dmt is a dummy variable that assumes a
value of one if year t equals the year in which individual m’s estab-
lishment closed down. 𝛼 m and 𝛿𝑡∗ (𝑚) represent individual, respectively,
displacement-year fixed effects and the vector Xmt contains a worker’s
14
The small dip in earnings of displaced workers a year before displacement is
common and usually attributed to early signs of distress in establishments that
are about to close.

131
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table 2
Effects of regional conditions on earnings losses upon displacement.

dep. var.: earnings increase (EUR) (1) (2) (3) (4)

∗∗∗ ∗∗∗ ∗∗∗
D -36.872 -37.872 -37.127 -90.431∗ ∗ ∗
(0.658) (1.857) (0.916) (19.135)
𝑀
𝐷 × 𝑂𝑖,𝑟 0.883 1.690 1.625
(1.613) (1.083) (1.076)
𝐻
𝐷 × 𝑂𝑖,𝑟 6.355∗ ∗ ∗ 6.666∗ ∗ ∗ 5.889∗ ∗ ∗
(1.537) (1.133) (1.126)
𝐷 × 𝐴𝑀
𝑖,𝑟
-0.102 -0.696 -0.165
(1.573) (1.036) (1.020)
𝐷 × 𝐴𝐻
𝑖,𝑟
-4.478∗ ∗ ∗ -4.925∗ ∗ ∗ -4.220∗ ∗ ∗
(1.667) (1.191) (1.170)
𝑀
𝑂𝑖,𝑟 0.438 0.463 0.558
(0.433) (0.569) (0.562)
𝐻
𝑂𝑖,𝑟 -0.040 -0.504 -0.076
(0.449) (0.736) (0.729)
𝐴𝑀
𝑖,𝑟
0.236 0.130 -0.203
(0.435) (0.617) (0.604)
𝐴𝐻
𝑖,𝑟
1.543∗ ∗ ∗ 1.454∗ 0.825
(0.463) (0.857) (0.837)
other interaction terms? no no no yes
age controls? yes yes yes yes
year dummies? yes yes
education dummies? yes yes yes yes
industry-year dummies? no no yes yes
region-year dummies? no no yes yes
R2 0.127 0.128 0.175 0.183
# obs. 90,688 90,688 90,688 90,688
∗∗∗
: p<.01, ∗ ∗ : p<.05, ∗ : p<.1. The dependent variable measures a worker’s change in
real daily earnings (in 2005 EUR), which is calculated as the (possibly zero) wage in the
year directly after the displacement event minus the wage in the last year in which the
worker is observed in the establishment that closes down. D is a displacement dummy
(1 for a displaced worker, 0 for a statistical twin). 𝑂𝑖𝑟𝑀 and 𝑂𝑖𝑟𝐻 form a dummy group
that captures whether the pre-displacement industry has a moderate (M) or high (H)
employment share in the region in which the worker was displaced. 𝐴𝑀 𝑖𝑟
and 𝐴𝐻𝑖𝑟
form
an analogous dummy group for the regional employment share of industries with a skill-
relatedness of 3 or higher to the pre-displacement industry. Age controls are the worker’s
age and squared age in the year of displacement. Education dummies group workers
into seven education classes. Industry dummies refer to the 5-digit industry and region
dummies to the labor market area in the displacement year. Both industry and region
dummies are interacted with displacement-year dummies. Standard errors are clustered
at the region-industry level.

age and age-squared in year t. ymt can be any of the following depen-
dent variables: daily earnings, the logarithm of daily wages, or a dummy
variable for the event that a worker is nonemployed, changes industries,
or changes regions. 𝑇𝑚𝑡 𝑘 is a dummy variable encoding event time. That
is, it takes the value one in observations that take place k years after the
(real or matched) displacement year (i.e., when 𝑡 = 𝑡∗ (𝑚) + 𝑘).
The parameters of interest are collected in vector 𝜏 2 . These point
estimates can be interpreted as the difference between displaced and
nondisplaced workers |k| years before or after the displacement event.
Fig. 1 graphs this vector, showing how the effects of displacement on
each of the dependent variables fade over time. First, note that pre-
displacement trends of displaced and nondisplaced workers are very
similar, suggesting that it is justifiable to interpret the effects depicted in
the figures as causal. Second, all of our dependent variables are strongly
affected by displacement, with most of the effects taking place in the
first year after displacement. Displacement reduces daily earnings by
about 37 EUR and keeps them depressed for the entire post-displacement
observation-window. Much of this reduction is due to the large drop in
employment rates, which reaches 38.4 percentage points (pp) in the first
post-displacement year. However, workers who get re-employed within
a year, face a fall in daily wages as well, of on average, 8.7%. Note that
these wage effects do not take into consideration any unemployment or
other benefits that displaced workers may receive. As a consequence, the
income effects of job-displacement will be less pronounced than the re-
duction in earnings reported here (see Schmieder et al. (2010) for a treat-
ment of unemployment benefits after job displacement in Germany).
Displacement also affects which jobs workers will choose. Displaced
workers are much more likely than their statistical twins to move out
of a labor market area (32.8 pp) or to change 5-digit industries (65.5
pp) right after they were displaced. Moreover, switching rates remain
elevated for at least two years after having been displaced. This suggests
that displaced workers do not immediately find well-matching jobs.
Local conditions as moderators of displacement effects
How does the local industry mix change the effect of displacement?
To study this, we interact the displacement dummy with information
on the employment shares of the pre-displacement and related indus-
tries in the region. However, these shares have strongly right-skewed
distributions. Therefore, we interact the displacement dummy with the
industry-mix dummy groups created in Eqs. (12) and (13), which are
robust to outliers. Ideally, we would integrate these interaction terms
in the difference-in-differences estimations of Eq. (15). However, given
that this would quintuple the number of parameters in the model, this
set-up would yield complex and hard-to-estimate interaction effects. In-
stead, we collapse the data to cross-sections (one for each displacement
year), where we observe workers at the time of displacement, t∗ . Next,
we pool the data from these three cross-sections to estimate models of

132
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Fig. 1. Difference-in-differences in post-displacement careers. Difference-in-differences estimates using Eq. (15), controlling for age, age-squared, education, year
and worker fixed effects. The dependent variables are daily earnings (in 2005 EUR, 1a), log(daily wage) (1b) and dummy variables for being nonemployed (1c),
switching regions (1d) and switching industries (1e). Region and industry switching are recorded in the last year in which a person worked in the job from which the
switch took place. As a consequence, switches recorded at 𝑡 = 𝑡∗ (𝑚) + 1 and 𝑡 = 𝑡∗ (𝑚) + 2 represent switches from one post-displacement job to another, not delayed
reemployment.

the following form:
𝑦𝑚𝑡∗ = 𝜅𝐷𝑚𝑡∗ + Π𝑖𝑟𝑡∗ −1 𝛾0 + 𝐷𝑚𝑡∗ Π𝑖𝑟𝑡∗ −1 𝛾1 + 𝑋𝑚𝑡∗ 𝛽 + 𝜂𝑖𝑡∗ + 𝜌𝑟𝑡∗ + 𝜖𝑚𝑡∗
where Π𝑖𝑟𝑡∗ −1 collects the dummy groups defined in the year be-
fore the displacement using Eqs. (12) and (13). 𝜂𝑖𝑡∗ and 𝜌𝑟𝑡∗ are
industry-displacement-year and region-displacement-year fixed effects
(for nondisplaced workers, these refer to the year in which their statisti-
cal twin was displaced). 𝑋𝑚𝑡∗ is a set of worker’s characteristics, includ-
(16) ing age, age-square, nationality dummies and a dummy group for the
worker’s educational attainment in the displacement year. The depen-
dent variable, 𝑦𝑚𝑡∗ , can be one of six variables: (1) worker m’s change
in earnings in the first year after displacement; (2) the change in daily

133
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table 3 Table 4
Effects of regional conditions on log(daily wage) upon displacement. Effects of regional conditions on short-term nonemployment.

dep. var.: log(wage gain) (1) (2) (3) (4) dep. var.: non-employed (y/n) (1) (2) (3) (4)
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
D -0.085 -0.097 -0.092 0.145 D 0.383 0.405 0.405 0.839∗ ∗ ∗
(0.005) (0.011) (0.009) (0.099) (0.006) (0.013) (0.008) (0.138)
𝑀 𝑀
𝐷 × 𝑂𝑖,𝑟 0.014 0.017 0.017 𝐷 × 𝑂𝑖,𝑟 -0.001 -0.015 -0.012
(0.012) (0.012) (0.012) (0.013) (0.009) (0.009)
𝐻 𝐻
𝐷 × 𝑂𝑖,𝑟 0.020∗ 0.023∗ ∗ 0.019∗ 𝐷 × 𝑂𝑖,𝑟 -0.055∗ ∗ ∗ -0.059∗ ∗ ∗ -0.055∗ ∗ ∗
(0.010) (0.009) (0.010) (0.015) (0.009) (0.009)
𝐷 × 𝐴𝑀
𝑖,𝑟
0.005 0.002 0.002 𝐷 × 𝐴𝑀
𝑖,𝑟
-0.017 -0.008 -0.008
(0.011) (0.010) (0.010) (0.015) (0.009) (0.009)
𝐷 × 𝐴𝐻
𝑖,𝑟
-0.006 -0.008 -0.010 𝐷 × 𝐴𝐻
𝑖,𝑟
0.010 0.006 0.008
(0.012) (0.012) (0.012) (0.013) (0.009) (0.009)
𝑀 𝑀
𝑂𝑖,𝑟 0.001 0.006 0.007 𝑂𝑖,𝑟 -0.006∗ ∗ -0.001 -0.002
(0.003) (0.004) (0.004) (0.003) (0.004) (0.004)
𝐻 𝐻
𝑂𝑖,𝑟 -0.001 0.008 0.010∗ 𝑂𝑖,𝑟 -0.002 0.009∗ 0.008
(0.003) (0.006) (0.006) (0.003) (0.005) (0.005)
𝐴𝑀
𝑖,𝑟
0.000 -0.002 -0.003 𝐴𝑀
𝑖,𝑟
-0.004 -0.001 -0.001
(0.003) (0.005) (0.005) (0.003) (0.005) (0.005)
𝐴𝐻
𝑖,𝑟
0.007∗ ∗ -0.003 -0.004 𝐴𝐻
𝑖,𝑟
-0.008∗ ∗ ∗ -0.007 -0.006
(0.003) (0.006) (0.006) (0.003) (0.006) (0.006)
other interaction terms? no no no yes other interaction terms? no no no yes
age controls? yes yes yes yes age controls? yes yes yes yes
year dummies? yes yes year dummies? yes yes
education dummies? yes yes yes yes education dummies? yes yes yes yes
industry-year dummies? no no yes yes industry-year dummies? no no yes yes
region-year dummies? no no yes yes region-year dummies? no no yes yes
R2 0.017 0.017 0.067 0.069 R2 0.210 0.212 0.260 0.266
# obs. 48,020 48,020 48,020 48,020 # obs. 90,688 90,688 90,688 90,688

Idem Table 2, with as dependent variable the change in log(daily wages) in the Idem Table 2, with as dependent variable a dummy for whether the worker was
first job after the displacement event. We only keep worker pairs for which both nonemployed in the year following the displacement event.
displaced worker and matched twin are employed in the year immediately after
displacement.

wages for workers who immediately find new jobs; a dummy variable
that indicates whether or not worker m remains nonemployed (3) for
one year or (4) for three years after displacement; (5) a dummy for
whether his first post-displacement job was in a different industry or
(6) in a different region than the job from which he was displaced.
The main parameters of interest - the interactions of local condi-
tions with the displacement dummy - are collected in 𝛾 1 and reported
in Tables 2 to 7. Each table refers to one of the dependent variables and
reports four different model specifications. The first column in these ta-
bles reports the overall effect of displacement, while controlling for a
worker’s age, education and nationality. These estimates should be sim-
ilar to the ones depicted in Fig. 1. The second column adds interactions
with local conditions. The third column adds industry-year and region-
year fixed effects. This is our preferred specification and most of the
discussion below will refer to this column. Finally, in the fourth col-
umn, we consider additional interactions of the displacement dummy
with a range of worker characteristics, as well as with a region’s size.
We will discuss these models at the end of this section.
Wages
Table 2 illustrates the adverse effects of displacement on earnings.
On average, workers lose about 37 EUR in daily earnings in the first
year after having been displaced (about 36% of their pre-displacement
earnings, see column 1). Table 4 shows that this is largely due to an
about 38 pp increase in nonemployment hazard. By contrast, for workers
who immediately find a new job, the loss in log(daily wages) is limited
to an 8.2% reduction (column 1, Table 3).15
As expected, these estimates are very close to the difference-in-
differences estimates in Fig. 1. However, effects vary with the local in-
dustry mix. Displacement-induced earnings-losses and nonemployment-
15
To keep the pairs balanced in these estimations, we require that not only the
displaced worker but also her statistical twin is employed in the year immedi-
ately after the displacement event.
risks are lower in locations with high employment shares of the pre-
displacement industry. Taking locations with low shares of the pre-
displacement and related industries as a benchmark and referring to
our preferred specification (column 3), the reduction in the earnings-
effect (see Table 2) due to having high employment shares of the old
industry amounts to 6.7 EUR (18%). This reduction is in part due to
changes in the effect on daily wages: for workers who find new jobs,
a large presence of the old industry in the region reduces the drop in
log(daily wage) by 0.023 log points (25%) (Table 3). Another part of
the reduction in earnings drop is due to lower rates of displacement-
related nonemployment. Having high instead of low local employment
shares of the old industry reduces the effect of displacement on short-
term nonemployment (Table 4) by 5.9 pp (or 15%) and on long-term
nonemployment rates by about 4.0 pp, a 21% reduction (Table 5).
The impact of skill-related employment in the region on wages and
nonemployment rates is somewhat different: it neither significantly re-
duces displacement-related nonemployment nor earnings losses. On the
contrary, high shares of related industries increase displacement-related
earnings losses by 4.9 EUR. At the same time, however, a large presence
of related industries protects workers from long-term nonemployment,
reducing the displacement effect by 1.1 pp.16 A potential explanation for
these findings is that jobs in related industries represent a lower quality
match compared to jobs in the pre-displacement industry. As a conse-
quence, skill-related employment opportunities in a region help workers
return to social-security covered employment sooner, but they do so at
the expense of the quality of the skill match.17
To study the effect of displacement on workers’ mobility, we drop
all displaced workers who disappear from the data for the entire 3-year
16
At 1.5 pp, the interaction effect with intermediate shares of related industries
is even higher, although this difference is not statistically significant.
17
In line with this, we find (not shown) that high local shares of related em-
ployment are associated with a reduced skill-relatedness between pre- and post-
displacement jobs.

134
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table 5 Table 6
Effects of regional conditions on long-term nonemployment. Effects of regional conditions on relocation upon displacement.
dep. var.: non-employed after 3 yrs (y/n) (1) (2) (3) (4) dep. var.: region switch (y/n) (1) (2) (3) (4)
∗∗∗ ∗∗∗ ∗∗∗
D 0.166 0.187 0.191 0.673∗ ∗ ∗ D 0.330∗ ∗ ∗ 0.364∗ ∗ ∗ 0.349∗ ∗ ∗ 0.208∗ ∗ ∗
(0.004) (0.008) (0.006) (0.189) (0.006) (0.013) (0.010) (0.069)
𝑀
𝐷 × 𝑂𝑖,𝑟 0.006 -0.003 -0.002 𝑀
𝐷 × 𝑂𝑖,𝑟 -0.049∗ ∗ ∗ -0.041∗ ∗ ∗ -0.041∗ ∗ ∗
(0.008) (0.006) (0.006) (0.013) (0.011) (0.011)
𝐻
𝐷 × 𝑂𝑖,𝑟 -0.034∗ ∗ ∗ -0.040∗ ∗ ∗ -0.037∗ ∗ ∗ 𝐻
𝐷 × 𝑂𝑖,𝑟 -0.034∗ ∗ -0.039∗ ∗ ∗ -0.044∗ ∗ ∗
(0.008) (0.006) (0.006) (0.014) (0.011) (0.011)
𝐷 × 𝐴𝑀
𝑖,𝑟
-0.019∗ ∗ -0.015∗ ∗ -0.016∗ ∗ 𝐷 × 𝐴𝑀 -0.008 -0.011 -0.015
𝑖,𝑟
(0.008) (0.006) (0.006) (0.015) (0.012) (0.012)
𝐷 × 𝐴𝐻
𝑖,𝑟
-0.011 -0.011∗ -0.010 𝐷 × 𝐴𝐻 -0.011 -0.009 -0.021∗
𝑖,𝑟
(0.008) (0.007) (0.007) (0.014) (0.012) (0.011)
𝑀
𝑂𝑖,𝑟 -0.005∗ ∗ 0.000 -0.000 𝑀
𝑂𝑖,𝑟 0.002 -0.001 -0.001
(0.002) (0.003) (0.003) (0.002) (0.004) (0.004)
𝐻
𝑂𝑖,𝑟 -0.003 0.009∗ ∗ 0.007∗ ∗ 𝐻
𝑂𝑖,𝑟 0.003 -0.009 -0.005
(0.002) (0.004) (0.004) (0.003) (0.006) (0.006)
𝐴𝑀
𝑖,𝑟
-0.002 -0.001 0.000 𝐴𝑀 -0.002 0.002 0.003
𝑖,𝑟
(0.002) (0.003) (0.003) (0.003) (0.005) (0.005)
𝐴𝐻
𝑖,𝑟
-0.005∗ ∗ 0.000 0.002 𝐴𝐻 -0.014∗ ∗ ∗ -0.008 -0.005
𝑖,𝑟
(0.002) (0.004) (0.004) (0.002) (0.007) (0.007)
other interaction terms? no no no yes other interaction terms? no no no yes
age controls? yes yes yes yes age controls? yes yes yes yes
year dummies? yes yes year dummies? yes yes
education dummies? yes yes yes yes education dummies? yes yes yes yes
industry-year dummies? no no yes yes industry-year dummies? no no yes yes
region-year dummies? no no yes yes region-year dummies? no no yes yes
R2 0.079 0.081 0.123 0.127 R2 0.176 0.178 0.242 0.246
# obs. 90,688 90,688 90,688 90,688 # obs. 71,108 71,108 71,108 71,108

Idem Table 2, with as dependent variable a dummy for whether the worker was Idem Table 2, with as dependent variable a dummy for whether a worker’s
nonemployed for at least three years after the displacement event. first post-displacement job is in a different labor market region than the pre-
displacement job. If a worker or his matched twin remains nonemployed, both
observations are dropped.

post-displacement observation window.18 For these workers, displace-

ment increases the likelihood of moving to another region by about 33
pp (column 1, Table 6) and of switching 5-digit industries by about 66
pp (column 1, Table 7).
The exact mobility choices, however, depend on the local industry
mix. Again, we use regions with low shares of the pre-displacement and
of related industries as a benchmark. Against this benchmark, we ob- Table 7
serve a 4.1 pp decrease in displacement-related region switching in re- Effects of regional conditions on switching industries upon displacement.
gions with a moderate employment share of the old industry (column 3
dep. var.: industry switch (y/n) (1) (2) (3) (4)
of Table 7) and a slightly lower (yet statistically indistinguishable) 3.9
∗∗∗ ∗∗∗ ∗∗∗
pp decrease in regions with a high share of employment in the old indus- D 0.657 0.718 0.687 0.885∗ ∗ ∗
(0.007) (0.012) (0.009) (0.069)
try. This is a modest change when compared to the 21 pp reduction in 𝑀
𝐷 × 𝑂𝑖,𝑟 -0.161∗ ∗ ∗ -0.134∗ ∗ ∗ -0.133∗ ∗ ∗
post-displacement industry-switching rates (Table 7) caused by the same (0.014) (0.010) (0.010)
variable. By contrast, high shares of related industries increase industry 𝐻
𝐷 × 𝑂𝑖,𝑟 -0.219∗ ∗ ∗ -0.207∗ ∗ ∗ -0.207∗ ∗ ∗
switching by 17 pp. These findings support our earlier conjecture that (0.014) (0.010) (0.010)
the presence of related industries helps workers find jobs faster in alter- 𝐷 × 𝐴𝑀
𝑖,𝑟
0.046∗ ∗ ∗ 0.067∗ ∗ ∗ 0.067∗ ∗ ∗
(0.015) (0.011) (0.011)
native industries, which represent relatively bad matches and therefore
𝐷 × 𝐴𝐻
𝑖,𝑟
0.152∗ ∗ ∗ 0.169∗ ∗ ∗ 0.167∗ ∗ ∗
pay somewhat lower wages. (0.015) (0.010) (0.010)
Overall, Tables 2 to 7 lead us to conclude that, whereas a presence 𝑀
𝑂𝑖,𝑟 -0.005 0.024∗ ∗ ∗ 0.024∗ ∗ ∗
of the old industry helps reduce displacement effects on earnings and (0.004) (0.005) (0.005)
𝐻
nonemployment, related industries only help displaced workers getting 𝑂𝑖,𝑟 -0.011∗ ∗ ∗ 0.026∗ ∗ ∗ 0.027∗ ∗ ∗
(0.004) (0.006) (0.006)
re-employed. However, a potential caveat is that, in spite of the match-
𝐴𝑀
𝑖,𝑟
0.012∗ ∗ ∗ -0.013∗ ∗ -0.013∗ ∗
ing design, workers may differ from one another in some unobserved (0.004) (0.006) (0.006)
(e.g., ability-related) characteristics. In that case, we would expect some 𝐴𝐻
𝑖,𝑟
0.010∗ ∗ ∗ -0.044∗ ∗ ∗ -0.043∗ ∗ ∗
sorting of workers across regions and industries based on these charac- (0.004) (0.007) (0.007)
teristics. It is therefore interesting that, although neither region nor in- other interaction terms? no no no yes
age controls? yes yes yes yes
dustry fixed effects were used in the matching procedure, adding them
year dummies? yes yes
(columns 3 of Tables 2 to 7) or not (columns 2) neither changes the point education dummies? yes yes yes yes
industry-year dummies? no no yes yes
region-year dummies? no no yes yes
18
As in the case of log(wage gain), we impose the same requirement on sta- R2 0.451 0.478 0.519 0.520
tistical twins (who may disappear due to attrition) to keep samples balanced. # obs. 71,102 71,102 71,102 71,102
Given that a worker’s willingness to change regions or industries may depend on
her likelihood of finding a new job, this design choice may lead to some sample Idem Table 2, with as dependent variable a dummy for whether a worker’s first
selection bias in the current analyses. However, this attrition does not affect the post-displacement job is in a different industry than the pre-displacement job.
tests in later multinomial logit models for the cross-over effects predicted by our If a worker or his matched twin remains nonemployed, both observations are
model. dropped.

135
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table 8 est earnings drop and lower nonemployment incidence we attributed

Multinomial post-displacement regression. to Marshallian externalities may instead be due to the specific type of
Outcome: workers that clusters attract. A similar problem occurs if our local in-
dustry groupings pick up differences in the size of the local economy. In
stay ind. & reg. switch reg. switch ind. switch ind. & reg.
that case, what matters is not the industry mix, but the total amount of
𝑀
𝑂𝑖,𝑟 2.265∗ ∗ ∗ 1.625∗ ∗ ∗ 0.876∗ ∗ ∗ 0.737∗ ∗ ∗ employment in the region. In essence, the effects would still be causal,
(0.193) (0.160) (0.042) (0.043)
𝐻
but the differences in these causal effects would arise from differences
𝑂𝑖,𝑟 3.544∗ ∗ ∗ 2.437∗ ∗ ∗ 1.010 0.918
(0.313) (0.280) (0.054) (0.063)
in, for instance, region-size or workers’ education, not local industry
𝐴𝑀 0.925 0.868 1.329∗ ∗ ∗ 1.371∗ ∗ ∗ mix.
𝑖,𝑟
(0.065) (0.090) (0.066) (0.088) Table A1 of the Appendix shows that different local conditions are
𝐴𝐻𝑖,𝑟 0.674∗ ∗ ∗ 0.511∗ ∗ ∗ 1.459∗ ∗ ∗ 1.412∗ ∗ ∗ indeed associated with different kinds of workers. Most saliently, loca-
(0.054) (0.056) (0.077) (0.091) tions where the pre-displacement and related industries have higher em-
log(reg. size) 1.027 0.981 1.007 0.899∗ ∗ ∗
(0.040) (0.065) (0.025) (0.030)
ployment shares tend to also have a higher educated workforce. To find
age controls? yes yes yes yes out whether this could explain the results presented above, we explore
education dummies? yes yes yes yes how much of the documented effect-heterogeneity can be attributed to
sector-year dummies? yes yes yes yes these worker characteristics (and to a region’s size). If our findings are
state-year dummies? yes yes yes yes
unaffected by accounting for these observable sources of heterogeneity,
log(L) -64,356 -64,356 -64,356 -64,356
# obs. 45,341 45,341 45,341 45,341 there is less cause for concern that unobservable sources of worker het-
# clust. 5,042 5,042 5,042 5,042 erogeneity drive our results. Therefore, we rerun the analyses of column
partial R2 0.049 0.049 0.049 0.049 3 of Tables 2 to 7, but now add interactions of the displacement dummy
∗∗∗
: p<.01, ∗ ∗ : p<.05, ∗ : p<.1. Multinomial regression of first job-switch within with a worker’s educational attainment, age and the logarithm of total
three years of displacement. Base category is composed of workers who do not employment in the region. Results on the interactions with local condi-
return to social-security covered jobs (nonemployment). Coefficients represent tions are reported in columns 4 of these tables. The estimated interaction
relative risk ratios. Standard errors, clustered at industry-region level, are re- effects of displacement with worker-level characteristics and region size
ported in parentheses. are reported in Table A2 of the Appendix.
Many of the new interaction effects are significant and interesting
in their own right. For instance, absolute earnings losses tend to in-
crease with educational attainment (column 1, Table A2). However,
Table 9
Multinomial cross-over effects.
this simply reflects that, for highly educated workers, earnings fall from
higher pre-displacement levels. Instead, differences in the relative drop
outcome: switch reg. outcome: switch ind. & reg. of daily wages (column 2) across education groups are barely statis-
base: stay ind. & reg. base: switch ind.
tically significant. However, point estimates suggest that the drop in
𝑀
𝑂𝑖,𝑟 0.717∗ ∗ ∗ 0.841∗ ∗ ∗ daily wages is relatively modest for workers with vocational training
(0.085) (0.047) (VT), high school and vocational training (HS+VT), or with a degree
𝐻
𝑂𝑖,𝑟 0.688∗ ∗ ∗ 0.909
from Germany’s - mostly vocational - technical colleges (C). This sug-
(0.091) (0.057)
𝐴𝑀 0.938 1.032
gests that what matters is how applied, not how long, workers’ education
𝑖,𝑟
(0.106) (0.060) is. Similar patterns emerge for the incidence of displacement-induced
𝐴𝐻
𝑖,𝑟
0.759∗ ∗ 0.968 short- and long-term nonemployment (columns 3 and 4), where voca-
(0.084) (0.059) tional training (VT and HS+VT) and degrees from technical colleges
Coefficients from Table 8, expressed against the base outcomes in the column are associated with shorter post-displacement nonemployment spells.
headers. Apparently, an applied education shields workers from some of the neg-
ative consequences of job displacement. Similarly, an applied education
is associated with lower post-displacement industry-switching rates. In
contrast, the degree to which displaced workers leave their region in-
estimates of displacement nor of the interaction effects much. However, creases monotonically with the level of education. Displacement effects
because the explanatory power of these fixed effects reduces the stan- furthermore change with age, although the statistical evidence for this
dard error of regression, adding them typically yields efficiency gains, is weaker. The size of a region is an important moderator as well: dou-
which manifest themselves in smaller standard errors. This is reassur- bling a region’s size cuts earnings losses by 2.9 EUR, daily wage losses
ing. After all, had there been any confounders, we would have expected by 1.1%, region-switching rates by 2 pp and industry-switching rates by
that they would exhibit at least some regional or industry variation. 1 pp.
The combination of an absence of notable shifts in point estimates and Overall, the findings in Table A2 imply substantial effect-
a tightening of confidence intervals suggests that the matching proce- heterogeneity across workers with different educational backgrounds
dure successfully removes any correlation between displacement and and ages. However, when comparing columns 4 to columns 3 in
confounders at the region or industry level. Therefore, the scope for Tables 2 to 7, adding these interactions barely changes how displace-
ability-related confounding by factors beyond (unobserved) region and ment effects vary with local conditions.19 This suggests that, although
industry characteristics would seem limited. displacement effects do depend on observable worker characteristics,
this dependence does not explain any of the moderating effects we
Robustness: effect heterogeneity have attributed to the local industry mix. We still cannot be certain
that the same holds for unobservable worker characteristics. However,
Unobserved worker characteristics may yet be problematic in a dif- given that important markers of individual productivity such as age
ferent way. So far, we have interpreted the heterogeneity in displace- and education do not seem to be part of the explanation, this would be
ment effects as evidence of Marshallian externalities. However, this remarkable.
effect-heterogeneity may not be driven as much by characteristics of
local industries as of the workers attracted to these industries. For in-
stance, firms in local clusters may attract more highly educated work- 19
Note that the main effect of displacement changes drastically in all tables.
ers than their peers outside those clusters. In that case, the more mod- However, this simply reflects a change in reference category.

136
F.M.H. Neffke et al.
6. Marshallian externalities and strategic search
A central prediction in search theory is that workers will search
harder when job prospects improve. Testing this prediction is difficult,
because search efforts are unobserved. After all, the fact that unem-
ployment spells are shorter when labor markets are tighter does not
necessarily imply greater search efforts in such episodes. Instead, the
reduction in unemployment duration could simply be due to an im-
provement in job arrival rates or wage offers. However, the model in
section 3 showed that the indirect effect of labor market conditions on
job-offer quality and arrival rates via search efforts can be isolated from
their direct effects by studying not just whether workers find new jobs,
but where they find these jobs. In particular, the model predicts that the
hazard of getting new jobs in industries other than their old industry
– holding labor market conditions in these other industries constant –
decreases when job prospects in the old industry improve. Finding such
effects would mean that workers strategically reallocate search efforts
toward the old industry. Fallick (1993) shows that these effects indeed
exist.20
We use this framework to explore whether workers also strategically
adjust their search efforts to Marshallian externalities. We do so by in-
terpreting what we have called “favorable local conditions” for a sector
as a large presence of this sector in the region. Such an interpretation
is in line with the literature on agglomeration externalities, which uses
spatial concentration indices to identify Marshallian externalities (e.g.,
Glaeser et al. 1992; Henderson et al. 1995). Moreover, we control for
a region’s size to make sure these effects are driven by the local labor
market’s composition, not just by its size.
In this context, the model of section 3 predicts that job searchers are
more likely to find a job in sectors that have a large local presence in the
region. This prediction derives from a combination of two effects: first,
suitable job offers will arrive at higher rates when local conditions in
a sector are favorable, which, second, induces workers to redirect their
search efforts toward this sector, raising arrival rates even further. To
illustrate this, imagine a region in which the pre-displacement sector
is relatively large. The Marshallian externalities in this region should
shift search efforts to the pre-displacement sector, away from the alter-
native, skill-related sector. As a consequence, holding local conditions
in related industries constant, the job-finding hazard in these related
industries should still drop. This is comparable to Fallick’s original pre-
diction: favorable conditions in the pre-displacement industry should
lead to a drop in the relative risk of finding a job in related industries
vis-à-vis staying nonemployed.21 Because a reduction in search efforts
will also limit the spatial scope of search, we derived a further predic-
tion, namely that a large local presence of the pre-displacement industry
will lead to a drop in the relative risk of finding nonlocal vis-à-vis local
jobs in related industries. The same predictions hold with the roles of
pre-displacement and related industries reversed.
To test these implications, we drop all nondisplaced workers and
keep only the sample of displaced workers. Presumably, all of these
workers have been confronted with an exogenous shock that requires
them to start searching for jobs, making them an ideal group in which
to test the predictions of our search model. To do so, we jointly estimate
20
Fallick does not use any spatial information in his tests. Moreover, his ev-
idence for strategic search is not robust across specifications but only emerges
when proxying labor market conditions in the old industry by the (national) em-
ployment growth in the industry, not when using other measures of industries’
success.
21
Note that this is not the same as a drop in the probability of finding jobs in
related industries. This probability will drop because more workers exit nonem-
ployment through jobs in the pre-displacement industry. However, the higher
job-finding rate in the pre-displacement industry will itself lower the likelihood
of staying nonemployed. From this, it is not clear how a local concentration of
jobs in the pre-displacement industry will affect the relative risk of accepting
jobs in related industries instead of remaining nonemployed.
137
Journal of Urban Economics 108 (2018) 124–140
how local conditions affect each of the potential search outcomes. That
is, we estimate the multinomial logit model proposed in section 3 with
five potential outcomes. The first outcome is that the worker does not
find a new job within three years after displacement. The other out-
comes are that the first job the worker finds is (2) in the same industry
and region, (3) in the same industry but in a different region, (4) in
a different industry but the same region or (5) in a different industry
and region than the job from which he was displaced. Table 8 reports
how local conditions affect relative risk ratios vis-à-vis the base category
of nonemployment. In this analysis, we control for age, age-squared,
log(region size) and education dummies. However, because of the non-
linearity of the multinomial logit model, we have to aggregate industry
and region dummies to the level of 15 broad sectors and the 16 German
states (Bundesländer) respectively.
We report the outcomes of these multinomial regressions in terms
of relative risk ratios. That is, parameter estimates reflect the increase
in the ratio of the likelihood that the outcome in each column-header
materializes instead of the base-line outcome of nonemployment, for
a one-unit increase in regressor value. Parameter estimates above one
indicate a positive association with this ratio, below one a negative as-
sociation. For instance, the parameter estimate of 1.41 for 𝐴𝐻 𝑖𝑟 in the
final column means that, when we compare workers in regions where
the employment share of related industries is high instead of low, the
ratio of the probability that a worker will find a job in a different re-
gion and industry to the probability that he will remain nonemployed
increases by a factor 1.41.
Higher local employment shares in a sector increase the likelihood
that workers find local jobs in that sector. Compared to the reference
category of regions with low employment shares of the old and of re-
lated industries, the relative risk of finding a local job in the old indus-
try vis-à-vis staying nonemployed is over twice (three times) as high
in regions with intermediate (high) employment shares of the old in-
dustry (first column of Table 8). Similarly, higher local employment
shares of related industries increase the relative risk of finding local jobs
outside the pre-displacement industry by factors of 1.3 and 1.5, respec-
tively (third column). These findings provide some first evidence that
Marshallian externalities directly affect offer arrival rates (and/or offer
quality).
Local conditions should also affect job-finding rates indirectly,
through the reallocation of search efforts. In line with this, we find that
high shares of related industries in the region decrease the likelihood
of finding local (first column) and nonlocal jobs (second column) in
the old industry compared to staying nonemployed. Similarly, the third
and fourth columns of Table 8 show that favorable conditions in the
old industry significantly decrease the relative risk of finding a new job
in other industries (be it local or nonlocal) vis-à-vis remaining nonem-
ployed, although we only find such effects when employment shares
in the pre-displacement industry are intermediate, not when they are
high.22
These indirect cross-over effects between the local conditions in one
sector and job-finding rates in the other sector are also visible when
looking at spatial aspects of job search. In particular, a large local
presence of one sector should reduce search efforts in the other sec-
tor, therewith limiting the spatial reach of search in this other sector.
Table 9 confirms this prediction. The table re-expresses the relative risk
ratios reported in Table 8 in a way that compares nonlocal to local job-
finding rates in the pre-displacement (first column) and other indus-
tries (second column). As predicted, high local shares of related indus-
tries decrease the relative risk of finding nonlocal instead of local jobs
22
Note that the effects on finding local jobs (third column) are smaller than
the effects on finding nonlocal jobs (fourth column). This aligns with our expec-
tations that the reduction of search efforts not only decreases job-finding rates,
but also shifts the geographical balance of job-finding rates towards local jobs.
Below, we will test this prediction formally.
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

in the old industry by 24.1% (first column). A similar effect is visible Our study can be extended in several ways. Our focus on Marshallian
for jobs outside the pre-displacement industry: a presence of the old in- externalities made it natural to study the role of local industry concentra-
dustry increases the relative risk of workers’ finding such jobs outside tions. However, workers’ human capital is not just specific to an indus-
instead of inside the region. However, here, we observe a statistically try, but also to occupations. It would therefore be interesting to explore
significant effect only for intermediate employment shares of the old the relative importance of geographical clusters of occupations instead
industry. of industries as studied by, for instance, Bleakley and Lin (2012). More-
In Online Appendix B, we show that the substance of these results over, national labor market institutions vary markedly across countries.
does not change when we add variables that describe the local condi- Consequently, displacement will have different consequences in differ-
tions in neighboring regions. Moreover, so far, we have calculated em- ent countries. Repeating the analyses of this paper in different regions of
ployment shares as shares of total employment reported in the social se- the world might therefore provide interesting lessons in how Marshal-
curity data. However, the amount of employment that is not covered by lian labor market externalities operate in different national contexts.
social security may differ by region. Therefore, as a robustness check we Finally, the finding that concentrations of the pre-displacement and
redefine local conditions based on shares that use a region’s population related industries help displaced workers find new jobs may have useful
as a denominator. This adjustment does not change the substance of our implications for economic policy. Currently, cluster-based policies and
outcomes either (see Online Appendix C). Finally, we explore whether local development programs like the European Union’s Smart Special-
any of our results are strongly driven by a particular time period or re- ization efforts often focus on innovation and the creation of new busi-
gion by splitting the sample of displaced workers into subsamples by nesses. However, our findings suggest that clusters also benefit workers,
displacement year and by the territories of former West and East Ger- offering alternative employment opportunities that protect against pro-
many. Results are reported in Online Appendix E. Given the smaller sizes tracted unemployment. Taking industries’ human capital requirements
of these subsamples, point estimates tend to be less precisely estimated. into account in cluster definitions could therefore increase the effective-
However, although there is some variation in the magnitude of effects ness of cluster policies.
across these subsamples, our main conclusions on the existence of Mar-
shallian externalities in job search, as well as their reflection in strategic Acknowledgments
search-effort allocation, find support in each of the subsamples. Taken
together, therefore, the findings in this subsection strongly support the The authors thank Juan Pablo Chauvin, Andres Gomez, Ricardo
notion that workers take Marshallian externalities into account when Hausmann, Ines Helm, Ljubica Nedelkoska, Stuart Russell, Sandhya
searching for jobs. Srinivas, Simon Wiederhold and the participants of the IWH Workshop
on Firm Exit and Job Displacement for valuable comments and Ingo
7. Conclusions Konradt for his excellent research assistance. Frank Neffke received fi-
nancial support from the MasterCard Center for Inclusive Growth. César
We have shown evidence for Marshallian externalities in how a re- Hidalgo received financial support from the MIT Media Lab consortia.
gion’s industry mix affects the post-displacement careers of workers
who lose their jobs in establishment closures. High concentrations of Appendix: Summary statistics
the pre-displacement industry reduce the earnings losses experienced
by these workers, predominantly by reducing the time it takes workers Table A1 provides summary statistics for worker-level characteristics
to find new jobs. In contrast, high concentrations of industries that are in region-industry combinations in the EH with different concentrations
related to the pre-displacement industry are associated with higher earn- of the old and related industries. Table A2 displays the interaction ef-
ings losses, but a lower long-term nonemployment incidence. In places fects of worker characteristics, as well as of a region’s size, with the
where these related industries are abundant, workers tend to find new displacement dummy for the models in columns (4) of Tables 2-7.
jobs sooner by opting to change industries. Furthermore, we find ev-
idence that suggests that workers take these Marshallian externalities Table A1
into consideration when allocating search efforts. Large concentrations Group averages of individual level characteristics.
of related industries not only reduce the relative risk of finding jobs in employment share old ind. employment share related ind.
the pre-displacement industry compared to remaining nonemployed. If
low medium high low medium high
workers still do find jobs in the pre-displacement industry, such concen-
trations also reduce the relative risk of finding nonlocal instead of local age 39.7 39.8 39.9 39.8 39.7 39.8
edu (ND) 11.27% 10.04% 9.49% 10.66% 10.72% 9.54%
jobs in that industry, showing that local concentrations of related in-
edu (VT) 65.66% 64.39% 66.53% 64.17% 65.98% 67.06%
dustries increase the spatial scope of job-search in the pre-displacement edu (HS) 0.54% 0.48% 0.41% 0.48% 0.49% 0.32%
industry. Similar cross-over effects are found in regions where the pre- edu (HS+VT) 2.64% 2.24% 2.66% 2.31% 2.69% 2.67%
displacement industry is large. edu (C) 1.97% 3.32% 3.75% 2.60% 3.13% 3.29%
These results are robust to a number of changes in the model speci- edu (U) 2.23% 3.26% 4.13% 2.62% 3.30% 3.83%
edu (miss.) 15.71% 16.27% 13.02% 17.16% 13.69% 13.29%
fication. For instance, adding the industrial composition of neighboring log(reg. size) 12.3 12.3 12.1 12.4 12.3 12.1
regions does not change any of the conclusions in the paper. Similarly,
controlling for industry and region fixed effects does not lead to any sig- Averages of a worker’s age and the natural log of a region’s total social-security-
nificant changes in point estimates. Furthermore, we explored whether covered employment, as well as shares of each education type (ND: no degree,
VT: vocational training, HS: high school, HS+VT: high school + vocational
our findings are driven by the sorting of workers across locations. Al-
training, C: (applied) college, U: University, miss.: missing educational informa-
though such sorting happens and although the characteristics of workers
tion) by group. Groups refer to categories based on the local employment share
moderate displacement effects, accounting for worker-level heterogene- of the old industry (the three left-most columns) or of industries related to the
ity in displacement effects does not alter the estimated effects of Mar- old industry (the three right-most columns).
shallian externalities.

138
F.M.H. Neffke et al. Journal of Urban Economics 108 (2018) 124–140

Table A2
Estimated interaction effects of individual level characteristics.

dependent variable:

earnings increase log(wage gain) nonemp. (short) nonemp. (long) reg. switch ind. switch

D -90.431∗ ∗ ∗ 0.145 0.839∗ ∗ ∗ 0.673∗ ∗ ∗ 0.208∗ ∗ ∗ 0.885∗ ∗ ∗

(19.135) (0.099) (0.138) (0.189) (0.069) (0.069)
D × log(reg. size) -2.950∗ ∗ ∗ -0.011∗ ∗ ∗ -0.001 0.006∗ ∗ -0.020∗ ∗ ∗ -0.009∗ ∗
(0.499) (0.004) (0.004) (0.003) (0.006) (0.005)
D × age -0.944∗ ∗ ∗ -0.003∗ ∗ ∗ 0.008∗ ∗ ∗ 0.004∗ ∗ ∗ 0.001 0.002∗ ∗ ∗
(0.058) (0.001) (0.000) (0.000) (0.000) (0.000)
D × age2 -0.0198∗ ∗ -0.0000 0.0004∗ ∗ ∗ 0.0001 -0.0001 0.0000
(0.0077) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
D × edu(VT) 2.863∗ ∗ ∗ 0.022∗ -0.104∗ ∗ ∗ -0.043∗ ∗ ∗ 0.017 -0.048∗ ∗ ∗
(0.900) (0.012) (0.010) (0.008) (0.011) (0.010)
D × edu(HS) -25.352∗ ∗ ∗ -0.003 0.035 0.111∗ ∗ ∗ 0.063 -0.000
(8.514) (0.057) (0.037) (0.041) (0.047) (0.041)
𝐷 × edu(HS+VT) -22.740∗ ∗ ∗ 0.030 -0.085∗ ∗ ∗ -0.004 0.135∗ ∗ ∗ -0.032
(3.613) (0.021) (0.019) (0.015) (0.022) (0.020)
D × edu(C) -13.132∗ ∗ ∗ 0.036 -0.157∗ ∗ ∗ -0.020 0.144∗ ∗ ∗ -0.099∗ ∗ ∗
(3.738) (0.023) (0.018) (0.015) (0.021) (0.026)
D × edu(U) -29.133∗ ∗ ∗ 0.015 -0.133∗ ∗ ∗ -0.010 0.162∗ ∗ ∗ -0.036∗
(5.088) (0.023) (0.017) (0.015) (0.022) (0.019)
D × edu(miss.) 5.138∗ ∗ ∗ 0.028∗ -0.093∗ ∗ ∗ -0.037∗ ∗ ∗ 0.036∗ ∗ ∗ -0.043∗ ∗ ∗
(1.140) (0.016) (0.012) (0.010) (0.013) (0.013)
age controls? yes yes yes yes yes yes
education dummies? yes yes yes yes yes yes
industry-year dummies? yes yes yes yes yes yes
region-year dummies? yes yes yes yes yes yes
R2 0.183 0.069 0.266 0.127 0.246 0.520
# obs. 90,688 48,020 90,688 90,688 71,108 71,102
∗∗∗
: p<.01, ∗ ∗ : p<.05, ∗ : p<.1. Estimated interaction effects of age, age-squared, education dummies and log(region size) with
the displacement dummy for models 4 in Tables 2–7. Age and log(region size) are expressed in deviation of their sample
means before creating interaction terms. The dependent variable for each column is indicated in the column headers.

Supplementary material
Supplementary material associated with this article can be found, in
the online version, at doi:10.1016/j.jue.2018.09.006.
References
Andersson, F., Haltiwanger, J.C., Kutzbach, M.J., Pollakowski, H.O., Weinberg, D.H.,
2014. Job displacement and the duration of joblessness: The role of spatial mismatch.
Technical Report. National Bureau of Economic Research.
Baptista, R., Costa, C., 2012. Agglomeration vs. Organizational Reproduction: The Molds
Cluster in Portugal. Technical Report. Papers in Evolutionary Economic Geography.
Becker, G.S., 1962. Investment in human capital: A theoretical analysis. The journal of
political economy 70 (5), 9–49.
Bender, S., Haas, A., Klose, C., 2000. Iab employment subsample 1975-1995: opportunities
for analysis provided by the anonymised subsample.
Black, S.E., Devereux, P.J., Salvanes, K.G., 2015. Losing heart? the effect of job displace-
ment on health. ILR Review. 0019793915586381.
Bleakley, H., Lin, J., 2012. Thick-market effects and churning in the labor market: Evi-
dence from us cities. Journal of urban economics 72 (2), 87–103.
Burda, M.C., Mertens, A., 2001. Estimating wage losses of displaced workers in germany.
Labour Economics 8 (1), 15–41.
Carrington, W.J., 1993. Wage losses for displaced workers: is it really the firm that mat-
ters? Journal of Human Resources 435–462.
Carrington, W. J., Fallick, B. C., 2015. Do we know why earnings fall with job displace-
ment?.
Couch, K.A., Placzek, D.W., 2010. Earnings losses of displaced workers revisited. The
American Economic Review 572–589.
Csáfordi, Z., Lőrincz, L., Lengyel, B., Kiss, K.M., 2018. Productivity spillovers through
labor flows: productivity gap, multinational experience and industry relatedness. The
Journal of Technology Transfer 1–36.
Dauth, W., 2010. The mysteries of the trade: employment effects of urban interindustry
spillovers. Technical Report. IAB discussion paper.
Davis, S.J., von Wachter, T.M., 2011. Recessions and the cost of job loss. Technical Report.
National Bureau of Economic Research.
Delgado, M., Porter, M.E., Stern, S., 2010. Clusters and entrepreneurship. Journal of Eco-
nomic Geography 10 (4), 495–518.
Diodato, D., Neffke, F., O’Clery, N., 2018. Why do industries coagglomerate? how mar-
shallian externalities differ by industry and have evolved over time. Journal of Urban
Economics.
Duranton, G., Puga, D., 2004. Micro-foundations of urban agglomeration economies.
Handbook of regional and urban economics 4, 2063–2117.
Eliason, M., Storrie, D., 2009. Does job loss shorten life? Journal of Human Resources 44
(2), 277–302.
Ellison, G., Glaeser, E.L., Kerr, W.R., 2010. What causes industry agglomeration? evidence
from coagglomeration patterns. American Economic Review 100, 1195–1213.
Eriksson, R.H., Henning, M., Otto, A., 2016. Industrial and geographical mobility of
workers during industry decline: The swedish and german shipbuilding industries
1970–2000. Geoforum 75, 87–98.
Fallick, B.C., 1992. Job security and job search in more than one labor market. Economic
Inquiry 30 (4), 742–745.
Fallick, B.C., 1993. The industrial mobility of displaced workers. Journal of Labor Eco-
nomics 302–323.
Feser, E.J., 2002. Tracing the sources of local external economies. Urban Studies 39 (13),
2485–2506.
Florida, R., Mellander, C., Stolarick, K., 2011. Geographies of scope: an empirical analysis
of entertainment, 1970–2000. Journal of Economic Geography lbq056.
Gathmann, C., Helm, I., Schönberg, U., 2014. Spillover effects in local labor markets:
Evidence from mass layoffs.
Gibbons, R., Katz, L.F., 1991. Layoffs and lemons. Journal of Labor Economics 351–380.
Glaeser, E.L., Kallal, H.D., Scheinkman, J.A., Shleifer, A., 1992. Growth in cities. Journal
of Political Economy 100 (6), 1126–1152.
Greenstone, M., Hornbeck, R., Moretti, E., 2010. Identifying agglomeration spillovers: Ev-
idence from winners and losers of large plant openings. Journal of Political Economy
118 (3), 536–598.
Helsley, R.W., Strange, W.C., 1990. Matching and agglomeration economies in a system
of cities. Regional Science and urban economics 20 (2), 189–212.
Henderson, V., Kuncoro, A., Turner, M., 1995. Industrial development in cities. Journal
of Political Economy 103 (5), 1067–1090.
Herberger, L., Becker, B., 1983. Sozialversicherungspflichtig beschäftigte in der
beschäftigtenstatistik und im mikrozensus. Wirtschaft und Statistik 4 (83), 290–304.
Hethey, T., Schmieder, J.F., 2010. Using worker flows in the analysis of establishment
turnover–evidence from german administrative data. FDZ Methodenreport 6 (en), 43.
Ho, D.E., Imai, K., King, G., Stuart, E.A., 2007. Matching as nonparametric preprocessing
for reducing model dependence in parametric causal inference. Political analysis 15
(3), 199–236.
Holm, J.R., Østergaard, C.R., Olesen, T.R., 2017. Destruction and reallocation of skills
following large company closures. Journal of Regional Science 57 (2), 245–265.
Howland, M., Peterson, G.E., 1988. Labor market conditions and the reemployment of
displaced workers. Industrial & Labor Relations Review 42 (1), 109–122.
Jacobson, L.S., LaLonde, R.J., Sullivan, D.G., 1993. Earnings losses of displaced workers.
The American Economic Review 685–709.
Jenkins, S.P., 2005. Survival analysis. Unpublished manuscript, Institute for Social and
Economic Research, University of Essex, Colchester, UK.
Kosfeld, R., Werner, D.-Ö.A., 2012. Deutsche arbeitsmarktregionen–neuabgrenzung nach
den kreisgebietsreformen 2007–2011. Raumforschung und Raumordnung 70 (1),
49–64.
Lazear, E.P., 1979. Why is there mandatory retirement? The Journal of Political Economy
1261–1284.

139
F.M.H. Neffke et al.
Marshall, A., 1890. Principles of economics. Macmillan London.
Mortensen, D.T., 1986. Job search and labor market analysis. Handbook of labor eco-
nomics 2, 849–919.
Nedelkoska, L., Neffke, F., Wiederhold, S., 2015. Skill mismatch and the costs of job dis-
placement.
Neffke, F., Henning, M., 2013. Skill relatedness and firm diversification. Strategic Man-
agement Journal 34 (3), 297–316.
Neffke, F.M., Henning, M., Boschma, R., 2012. The impact of aging and technological
relatedness on agglomeration externalities: a survival analysis. Journal of Economic
Geography 12 (2).
Neffke, F.M., Otto, A., Weyh, A., 2017. Inter-industry labor flows. Journal of Economic
Behavior & Organization 142, 275–292.
140
Journal of Urban Economics 108 (2018) 124–140
Porter, M., 2003. The economic performance of regions. Regional studies 37 (6-7),
545–546.
Rosenthal, S.S., Strange, W.C., 2001. The determinants of agglomeration. Journal of urban
economics 50 (2), 191–229.
Schmieder, J.F., von Wachter, T., Bender, S., 2010. The long-term impact of job displace-
ment in Germany during the 1982 recession on earnings, income, and employment.
Technical Report. IAB discussion paper.
Schwerdt, G., 2011. Labor turnover before plant closure: Leaving the sinking ship vs cap-
tain throwing ballast overboard. Labour Economics 18 (1), 93–101.
Timmermans, B., Boschma, R., 2013. The effect of intra-and inter-regional labour mobility
on plant performance in denmark: the significance of related labour inflows. Journal
of Economic Geography lbs059.